JP4132766B2

JP4132766B2 - Image processing apparatus and method

Info

Publication number: JP4132766B2
Application number: JP2001307568A
Authority: JP
Inventors: 明村形
Original assignee: Ricoh Co Ltd
Current assignee: Ricoh Co Ltd
Priority date: 2001-10-03
Filing date: 2001-10-03
Publication date: 2008-08-13
Anticipated expiration: 2021-10-03
Also published as: JP2003115031A

Description

【０００１】
【発明の属する技術分野】
本発明は、画像処理装置および方法に関し、特に、文字認識処理のための文字画像二値化処理を行う画像処理装置および方法に関する。
【０００２】
【従来の技術】
文書のデジタル化が進み、紙面上に形成されている文字情報をデジタル化して画像情報として保存することが多くなった。しかし、ビットマップなどの形式の画像情報ではデータ量が大きく、さらに文書（画像情報）の編集が困難なため、文字コードに変換することが望まれている。
【０００３】
そこで、文字画像を読み取って文字コードに変換するＯＣＲ（Optical Character Reader：光学式文字読み取り装置）が利用されるが、ＯＣＲによって文字コードに変換する処理を行う際には、文字画像ごとに切り出して、その１文字画像を正規化して、文字境界領域のエッジの方向をもとに認識するので、スキャナなどの入力装置でデジタル化した多値の文字画像を、まず、文字画素の位置を特定するために文字画素と背景画素に分離する処理が必要となる。
【０００４】
スキャナ入力によって多値化された文字画像を入力として二値化処理し、二値表現された文字画像を出力する処理が必要となる。一般的な文字画像二値化処理を行う装置の構成を図２７に示す。この装置は、紙面上に形成されている文字情報をデジタル化して多値表現の画像データとして読み取るスキャナ部１、スキャナ部１で読み込んだ多値画像データや画像処理後の画像データを保存する画像記憶部２、入力された多値画像データを二値化処理するＣＰＵ（信号処理部）３、二値表現されている文字画像データをもとに文字認識処理を行うＯＣＲ４で構成される。また、図２７の矢印は、画像情報の流れを示している。
【０００５】
処理の流れは以下のようなものである。スキャナ部１で読み込んだ多値文字画像を、一旦、画像記憶部２に保存する。次に、その多値文字画像に対し、ＣＰＵ３において、スキャナ入力によって生じる文字の境界領域を分離する処理を行い、文字境界領域の分離画像を画像記憶部２に保存する。画像記憶部２に保存している入力多値文字画像と、ＣＰＵ３において画像処理して作成した文字境界領域の分離画像とをもとに、ＣＰＵ３において入力多値文字画像に対して、順に、二値化処理と、文字境界領域の平滑化処理と、を行う。この二値化処理および文字境界領域平滑化処理の結果、得られる二値文字画像データを、再び画像記憶部２に保存する。画像記憶部２に二値文字画像を保存した後、ＯＣＲ４は、この二値文字画像を入力として文字認識処理を行う。ＯＣＲ４内では、単純閾値で二値化するようになっているので、文字領域と判断された画素を濃度値２５５、背景画素と判断された画素を濃度値０とした８ｂｉｔの画像を文字認識の対象画像としてＯＣＲ４に送る。
【０００６】
文字や文書画像を二値化処理する従来技術として、特開２０００−３３３０２２号公報は、画像の二値化方法および装置並びに記憶媒体について開示している。また、特開平５−２８２４９４号公報は、画像データの二値化装置について開示している。また、特開平１０−３０８８７１号公報は、二値化方法および文字読み取り装置について開示している。これらの従来技術は、文字や文書画像を対象とした二値化方法および装置であるが、二値化処理を行う際の情報として、濃度情報を用いて行っている。
【０００７】
そのために、入力装置、特にスキャナで読み取った文字画像は、スキャナ特性によって、元が二値の文字情報でも、ＣＣＤの読み取りやＭＴＦ特性によって多値化し、特に、濃度勾配が急な文字境界領域で濃度断面がなまる。スキャナ特性は個々のスキャナ装置で固有のものであり、濃度断面のなまる度合いはスキャナごとに異なるため、多値化された文字画像では、濃度値のエッジの傾きが変化する。特に、なまる度合いが急なスキャナ入力の多値化画像の濃度断面部では、中間濃度になりやすい。二値化処理の際、その中間濃度領域に分布する画素の二値化判断が難しく、従来の二値化手法では、この境界領域の二値化判断を誤ることにより文字画像が潰れたり掠れたりするという問題がある。
【０００８】
また、濃度断面のなまる度合いは、局所的な領域情報、例えば、太い文字線に囲まれた領域は、背景領域であっても濃度値が近傍の文字領域の影響を受けて高くなるので、大きくなる傾向がある。スキャナ入力によって濃度断面がなまり、濃度断面がなまった結果生じる領域は主に文字境界領域で、文字画素から背景画素へまたは背景画素から文字画素へと遷移する領域なので、文字画素と背景画素が混在する。
【０００９】
また、特に、低解像度の影響でサンプリングが荒くなり、デジタル化する際に特に濃度変化の大きい文字境界領域において濃度値のばらつきが生じやすくなる。図２８に、解像度の異なる多値画像の二値化処理において、多値画像の解像度の違いにより二値化処理後の画像に文字境界領域において凹凸が生じる様子を示す。左側は、スキャナによって高解像度で取り込まれた画像を示す。右側は、スキャナによって低解像度で取り込まれた画像を示す。特に、右側上図の低解像度画像では、サンプリング間隔が大きいので、文字境界領域の急な濃度勾配を表現できず、濃度のばらつきが生じている。これらの画像を単純閾値（濃度値１２８、※１画素８ｂｉｔ表現）で二値化処理した結果が下段の図である。このように、特に低解像度の画像では、文字境界領域の急な濃度勾配を表現できないことから、従来の二値化手法での二値化後の画像では凹凸が生じやすい。
【００１０】
スキャナ入力した多値文字画像に対し濃度情報のみを用いて二値化すると、前述したように、スキャナ特性によって生じる濃度断面のなまる度合いが局所的に異なるので、二値化処理の結果、文字が潰れたり掠れたりし、ＯＣＲ認識率が低下するという問題があった。従来の二値化手法では、スキャナ入力によって生じる濃度断面がなまり、濃度断面がなまった結果生じる領域で二値化判定を誤っていることが多かった。濃度断面がなまった結果生じる領域に対して誤判定しにくい二値化処理方法および装置がＯＣＲ認識率向上のために望まれている。
【００１１】
また、前述したように低解像度入力によって文字境界領域で凹凸が表れやすくなり、ＯＣＲ認識率が低下するという問題があった。ＯＣＲは、文字画像の境界領域の文字の境界の方向を検出して認識を行う。ＯＣＲ認識率を向上させるために、認識するための特徴量にしている、従来方法では二値化処理の誤判定が多い、文字境界領域の情報を忠実に二値化することのできる方法および装置が求められている。
【００１２】
【発明が解決しようとする課題】
本発明は、かかる問題点に鑑みてなされたものであり、ＯＣＲでの文字認識処理のために、多値文字画像から二値文字画像を出力する二値化処理において文字境界領域をできる限り忠実に二値化し、ＯＣＲ認識率を向上させることのできる画像処理方法および装置を提供することを目的とする。
【００１３】
請求項１記載の発明は、スキャナ入力によって生じる文字境界領域に対し、文字の局所的な情報である濃度断面の情報を二値化判定に用いることによって、文字境界領域の二値化誤判定を抑え、ＯＣＲ認識率を向上させることを目的とする。
【００１４】
文字画像はコントラストが高いので、特にスキャナ入力の文字画像は文字の境界領域が劣化してなまりやすくなる。スキャナ入力によって生じる文字の境界領域は、文字線の局所的な位置情報やフォントの太さによっても文字の境界領域が劣化の度合いが近傍領域の濃度情報に影響を受けて変化するので、特に全画素に対し単一に濃度値の閾値を設定する従来の方法では二値化後の文字の潰れや掠れが生じていた。請求項２記載の発明は、文字の境界領域の二値化誤判定を防止するため、スキャナ入力によって生じる文字の境界領域を抽出し、この領域に適した二値化を行い、ＯＣＲの認識率を向上させることを目的とする。
【００１５】
文字画像はコントラストが高いので、特にスキャナ入力の文字画像は文字の境界領域が劣化してなまりやすくなる。スキャナ入力によって生じる文字の境界領域は、文字線の局所的な位置情報やフォントの太さによっても変わるので、特に、特に全画素に対し単一に濃度値の閾値を設定する従来の方法では二値化後の文字の潰れや掠れが生じていた。請求項３記載の発明は、二値化の判定に困難な画素に対しＬＢＧアルゴリズムを用いることによって、効率良く分離を行って、このスキャナ入力によって生じる文字境界領域を抽出し、この領域に適した二値化処理を行い、ＯＣＲの認識率を向上させることを目的とする。
【００１６】
スキャナ入力によって生じる文字境界領域は、文字画素と背景画素が混在する。請求項４記載の発明は、その領域を二値化処理するために、１濃度断面ごとに局所的な二値化閾値を設けて１濃度断面ごとに二値化処理を行い、局所的な二値化閾値の設定によって、二値化後の文字画像の潰れや掠れを防ぎ、ＯＣＲの認識率を向上させることを目的とする。
【００１７】
請求項５記載の発明は、スキャナ入力によって生じる文字境界領域が二値化された画像に対し、文字の内側の領域と背景領域とを近傍の二値化結果を用いて安定して二値化処理することを目的とする。
【００１８】
ＯＣＲで文字認識を行いたい場合、特に、低解像度の影響により、二値化処理後に文字境界領域に生じる凹凸によって誤認識することがある。請求項６記載の発明は、二値化後の文字画像に対し、境界領域に生じる凹凸を文字境界領域の平滑化で除去することによって、ＯＣＲの認識率を向上させることを目的とする。
【００１９】
スキャナ入力によって生じる文字の境界領域は、スキャナ特性や文字フォントの影響により文字境界領域が劣化してなまりやすくなる。スキャナなどで読み込んだ文字画像は、文字近傍領域の影響も受けるので、文字領域や背景領域でも濃度値の値が大きく異なる。スキャナ入力の画像に対して、全画素に単一の濃度の閾値を設定して二値化すると、文字が潰れや掠れの原因になる。請求項７記載の発明は、検知手段より、スキャナ入力によって生じる文字境界領域を分離して領域ごとに二値化処理を切り替えることによって、二値化後の文字境界電子領域での二値化誤判定を抑え、ＯＣＲの認識率を向上させることを目的とする。
【００２０】
スキャナ入力によって生じる文字境界領域は、文字画素と背景画素が混在する。請求項８記載の発明は、その領域を二値化するために、１断面ごとに局所的な二値化閾値を設けて１濃度断面ごとに二値化処理を行う。濃度断面の形状を用いて局所的な二値化閾値の設定を行うことによって、二値化後の文字画像の潰れや掠れを防ぎ、ＯＣＲの認識率を向上させることを目的とする。
【００２１】
【課題を解決するための手段】
かかる目的を達成するために、請求項１記載の発明は、紙面上に形成されている文字情報をデジタル化して多値画像情報として取り込むスキャナ手段と、画像処理演算を行う演算処理手段と、画像情報と演算処理手段による画像処理結果とを格納する画像記憶手段と、を備える画像処理装置であって、スキャナ手段による入力によって生じる文字境界領域を分離する分離手段と、文字境界領域を二値化処理する第１の二値化手段と、文字境界領域以外の領域を二値化処理する第２の二値化手段と、二値化処理後に文字境界領域の凹凸を平滑化して補正する補正手段とを有する画像処理装置であって、第１の二値化手段は、文字境界領域に関して水平、垂直、±４５度方向毎に注目画素から確定画素までに分布する非確定画素を近傍領域として、その近傍領域の平均濃度値に応じて設定された閾値に基づいて文字と背景とに二値化処理し、水平、垂直、±４５度方向毎の前記二値化結果に対し、近傍領域の濃度値と文字画像の濃度勾配に基づく一次微分値である抽出結果を重み付けして、二値化結果が文字であるすべての方向の一次微分値の合計値と二値化結果が背景であるすべての方向の一次微分値の合計値とを比較して大きいほうを最終二値化結果としての文字または背景として決めるために投票処理を行うことを特徴としている。
【００２２】
請求項２記載の発明は、請求項１記載の発明において、分離手段は、一次微分値と濃度値の情報を特徴量として用いて、文字境界領域の分離を行うことを特徴としている。
【００２３】
請求項３記載の発明は、請求項１記載の発明において、分離手段は、入力画像の一次微分−濃度平面を用いて、ＬＢＧアルゴリズムにより文字境界領域の分離を行うことを特徴としている。
【００２４】
請求項４記載の発明は、請求項１記載の発明において、第１の二値化手段は、文字境界領域を、スキャナ手段による入力によって生じる文字濃度断面情報を用いて二値化処理することを特徴としている。
【００２５】
請求項５記載の発明は、請求項１記載の発明において、第２の二値化手段は、近傍の文字境界領域の二値化処理結果を用いて二値化処理を行うことを特徴としている。
【００２６】
請求項６記載の発明は、請求項１記載の発明において、補正手段は、二値化処理後の画像の文字境界領域に生じる凹凸を検出して埋めることにより平滑化を行うことを特徴としている。
【００２７】
請求項７記載の発明は、請求項１記載の発明において、文字境界領域を分離して、分離された領域ごとに二値化処理を切り替えることを特徴としている。
【００２８】
請求項８記載の発明は、請求項１記載の発明において、第１の二値化手段は、文字境界領域を、スキャナ手段による入力によって生じる文字濃度断面とその形状の情報とを用いて二値化処理することを特徴としている。
【００２９】
請求項９記載の発明は、スキャナ入力の多値文字画像について、スキャナ入力によって生じる文字境界領域を分離する分離ステップと、分離された文字境界領域を二値化処理する第１の二値化ステップと、文字境界領域以外の領域を二値化処理する第２の二値化ステップと、二値化処理後に文字境界領域の凹凸を平滑化して補正する補正ステップとを有する画像処理方法であって、第１の二値化ステップでは、文字境界領域に関して水平、垂直、±４５度方向毎に注目画素から確定画素までに分布する非確定画素を近傍領域として、その近傍領域の平均濃度値に応じて設定された閾値に基づいて文字と背景とに二値化処理し、水平、垂直、±４５度方向毎の二値化結果に対し、近傍領域の濃度値と文字画像の濃度勾配に基づく一次微分値である抽出結果を重み付けして、二値化結果が文字であるすべての方向の一次微分値の合計値と二値化結果が背景であるすべての方向の一次微分値の合計値とを比較して大きいほうを最終二値化結果としての文字または背景として決めるために投票処理を行うことを特徴としている。
【００３０】
請求項１０記載の発明は、請求項９記載の発明において、分離ステップは、一次微分値と濃度値の情報を特徴量として用いて、文字境界領域の分離を行うことを特徴としている。
【００３１】
請求項１１記載の発明は、請求項９記載の発明において、分離ステップは、入力画像の一次微分−濃度平面を用いて、ＬＢＧアルゴリズムにより文字境界領域の分離を行うことを特徴としている。
【００３２】
請求項１２記載の発明は、請求項９記載の発明において、第１の二値化ステップは、文字境界領域を、スキャナ入力によって生じる文字濃度断面情報を用いて二値化処理することを特徴としている。
【００３３】
請求項１３記載の発明は、請求項９記載の発明において、第２の二値化ステップは、近傍の文字境界領域の二値化処理結果を用いて二値化処理を行うことを特徴としている。
【００３４】
請求項１４記載の発明は、請求項９記載の発明において、補正ステップは、二値化処理後の画像の文字境界領域に生じる凹凸を検出して埋めることにより補正を行うことを特徴としている。
【００３５】
請求項１５記載の発明は、請求項９記載の発明において、文字境界領域を分離して、分離された領域ごとに二値化処理を切り替えることを特徴としている。
【００３６】
請求項１６記載の発明は、請求項９記載の発明において、第１の二値化ステップは、文字境界領域を、スキャナ入力によって生じる文字濃度断面とその形状の情報とを用いて二値化処理することを特徴としている。
【００３７】
【発明の実施の形態】
以下、本発明の実施の形態を添付図面を参照しながら詳細に説明する。
【００３８】
［第１の実施例］
図１は、本発明の第１の実施の形態における画像処理方法を示す図である。この画像処理は、ＣＰＵ（信号処理部）が行う処理である。本画像処理方法は、順に、入力である多値文字画像Ｉ１に対しての文字境界領域の抽出処理１０１、多値文字画像Ｉ１と文字境界領域の抽出処理後に得られる文字境界領域の分離画像Ｉ２とを元にした文字境界領域の二値化処理１０２、文字境界領域の二値化処理後の画像に対する文字境界領域以外の領域の二値化処理１０３、および、上記二値化処理後の出力画像に対する文字境界領域の補正処理１０４を行い、二値文字画像Ｉ３を出力するものである。
【００３９】
文字境界領域の抽出処理１０１では、スキャナ入力によって生じる文字の境界領域を、（濃度値曲線の）一次微分などの特徴量を元に分離する。
【００４０】
文字境界領域の二値化処理１０２では、スキャナ入力によって生じる文字の境界領域に注目し、局所的に判定用閾値を変化させて二値化処理を行う。図１１に、スキャナから取り込んだ明朝体とゴシック体の文字画像の例とその二値化結果の例を示す。図１１において、上部の画像は原画像を示す。中央の上下の図は、原画像の濃度断面上の濃度分布を示している。下部の画像は、濃度分布上に示した閾値Ａと閾値Ｂそれぞれで二値化処理した結果の画像を示している。この処理では、図１１のように、対象文字画像の横方向にラスタ走査を行い、濃度断面を横方向ごとに設定する。文字境界領域であるエッジ部は、前述した文字境界領域の抽出処理１０１によって分離した結果の画像を用いている。
【００４１】
中央図の閾値Ａは、８ｂｉｔ画像の中間の濃度値である１２８を単純閾値としたものである。それに対し、閾値Ｂは、文字境界領域の抽出処理１０１において、文字境界領域であるエッジ部と判断された領域（中央図上段の点線部）に、その最大濃度画素と最小濃度画素の中間値に設定されたものである。その他の文字境界領域であるエッジ部に対しても、同様に、閾値Ｂを設定する。図１１下部画像は、以上の処理を横方向に全走査した後の二値化処理結果である。図１１下部画像を参照してわかるように、全体の画像から単純閾値Ａを設定して二値化を行うと、スキャナ入力によって生じる文字境界領域に対し、文字線が細い明朝体では掠れが生じ、文字線が太いゴシック体では潰れが生じている。それに対し、１濃度断面（エッジ部）ごとに局所的な閾値Ｂを設定した方法では、全濃度断面の山と谷を再現できるので、文字の掠れや潰れが生じにくくなっている。
【００４２】
なお、実際にはこの二値化処理１０２では、文字境界領域であるエッジ部のみを二値化するが、従来の単純閾値による手法と比較するために、文字境界領域以外の画素は閾値Ａで二値化した例を示している。
【００４３】
文字境界領域以外の領域の二値化処理１０３では、上記１０２の処理においてスキャナ入力によって生じる文字境界領域が二値化されているので、文字線の内部領域と背景領域とを、その近傍の二値化されたスキャナ入力によって生じる文字境界領域の情報を利用し統合させることによって、安定して二値化処理する。
【００４４】
文字境界領域の補正処理１０４では、特に低解像度の影響でサンプリングが荒くなり、デジタル化する際に特に濃度変化の大きい文字境界領域で、濃度値のばらつきが生じやすくなる文字境界領域に対しての補正処理を行う。低解像度の画像では文字境界領域の急な濃度勾配を表現できないことから二値化後画像で凹凸が生じやすいので、二値化処理後の文字画像に対し、境界領域に生じる凹凸を文字境界領域の平滑化処理で除去する。
【００４５】
図２は、図１に示される本発明の画像処理方法の基本構成を基にした二値化処理の流れを示す図である。図１における文字境界領域抽出処理１０１に対し、図２の処理では、非確定画素の分離処理２０１を行う。入力画像として、スキャナ部で取り込んだ多値文字画像に対し、文字境界領域において、近傍の濃度変化が急で、スキャナ入力によって生じる文字境界領域に分布する画素を非確定画素Ｐ１とし、文字境界領域以外において、近傍の濃度変化が平坦な領域に分布する画素を確定画素Ｐ２として分離処理を行う。
【００４６】
図１２に、非確定画素の分離処理２０１を施した画像の図を示す。左側の図はスキャナ入力によって取り込んだ多値文字画像の例であり、その原画像に対し、非確定画素の分離処理２０１を施した画像が右側の図である。非確定画素Ｐ１は白で、確定画素Ｐ２は黒で示されている。右図のように、非確定画素Ｐ１はスキャナ入力によって生じる文字境界領域なので、スキャナ特性によって生じる文字の滲んだ境界領域に分布している。確定画素Ｐ２は、スキャナ入力によって生じる文字境界領域以外の領域なので、文字線の内側の領域と、文字の影響がない背景領域とに分布している。主に文字境界領域に分布する画素を非確定画素Ｐ１と定義する。文字の影響がない背景領域に分布する画素と文字線の内側の領域を確定画素Ｐ２と定義する。具体的な分離方法として、非確定画素Ｐ１は近傍の濃度変化が急なことを利用して、濃度の一次微分特性を特徴量として分離処理を行う。入力の多値文字画像に対してエッジ検出フィルタをかけて一次微分の強度を算出する。この一次微分の強度に閾値を設けて閾値以上の画素を非確定画素Ｐ１、閾値未満の画素を確定画素Ｐ２として文字境界領域の分離画像２０８を得る。
【００４７】
次に、文字境界領域の二値化処理１０２として、非確定画素Ｐ１の二値化処理２０２を行う。入力は、スキャナ部で取り込んだ多値文字画像と文字境界領域の分離画像である。多値文字画像は、濃度情報から閾値決定を行うために用いる。文字境界領域の分離画像は、対象画像が、文字境界領域に分布する非確定画素Ｐ１か否かを調べるために用いる。注目している非確定画素を中心に局所的な正方形の近傍領域を設定し、その近傍領域内に分布する画素の平均濃度値を、注目している非確定画素の閾値として用いて二値化処理を行う。
【００４８】
次に、文字境界領域以外の領域の二値化処理１０３として、確定画素Ｐ２の二値化処理２０３を行う。入力として、非確定画素Ｐ１の二値化処理２０２後の、スキャナ入力によって生じる文字境界領域のみが二値化された画像を用いる。この非確定画素Ｐ１が分布する文字境界領域の周辺領域を二値化するために、注目する確定画素の近傍に分布する非確定画素Ｐ１の情報をもとに、確定画素Ｐ２の近傍には文字画素が多いか背景画素が多いか計算して判定を行い、多い方に統合する。
【００４９】
次に、第１の実施例における画像処理方法の文字境界領域の補正処理１０４に対して、文字境界領域の補正処理２０４を行う。入力は、確定画素Ｐ２の二値化処理２０３において全画素二値化された画像を用いる。図１３に示すような、Ｘ方向（横）とＹ方向（縦）の左上の画素からラスタ走査を行って横と縦方向の文字画素の連続領域を検出し、その連続領域の間が狭い場合、その画素を文字画素に埋めることによって横と縦方向の文字境界の凹凸を埋める。図１３では、２と４の例で、黒画素（文字画素）の連続の間に生じている白画素（背景画素）を黒画素（文字画素）に補正する。
【００５０】
次に、図１３に示すようなＸ方向とＹ方向の左上の画素から、横と縦方向の背景画素の連続領域を検出し、その連続領域の間が狭い場合、その画素を背景画素に埋めることによって、横と縦方向の文字境界の凹凸を埋める。図１３では、１と３の例で、白画素（背景画素）の連続の間に生じている黒画素（文字画素）を白画素（背景画素）に補正する。文字画像の境界領域の凹凸を埋めて文字境界領域を補正した二値文字画像を出力する。
【００５１】
また、図３に、図２の画像処理方法とは異なるバリエーションの二値化処理の流れを示す。図２の方法との違いは、確定画素Ｐ２の二値化処理３０３において、入力である多値文字画像Ｉ１と、非確定画素Ｐ１のみの二値化処理画像と、を利用して確定画素Ｐ２を二値化する処理を行う。確定画素Ｐ２は、濃度値が十分高いか十分低いかに分離できるので、図２の手法のように近傍の情報を利用しないで、濃度値に閾値を設けて確定画素Ｐ２の二値化を行う。その他の処理は、図２の方法と同様である。
【００５２】
［第２の実施例］
図４に、第１の実施例で説明した文字境界領域の抽出処理１０１に対し、本発明の第２の実施例における画像処理方法での非確定画素Ｐ１の分離処理を示す。この非確定画素Ｐ１の分離処理では、文字画像の濃度断面から領域を仮定し、多値文字画像Ｉ１を入力として、スキャナ入力によって生じる文字境界領域に分布する非確定画素Ｐ１を抽出する。
【００５３】
図１４に文字画像の濃度断面を示す。領域１は、濃度勾配が平坦な背景領域を、領域２は、文字から背景へ／背景から文字への遷移領域を、領域３は、濃度勾配が平坦な文字領域を、領域４は、文字線と文字線に挟まれた狭い背景領域への遷移領域をそれぞれ示している。図１４のような領域があるとまず仮定し、非確定画素Ｐ１が分布する領域２および４を一次微分の強度と濃度情報とをもとに抽出する。一次微分の強度は、一次微分量の算出処理４０１を行って得る。
【００５４】
図１５に一次微分と濃度値の分布を示す。図１４のそれぞれの領域１〜４は、一次微分−濃度値平面で見ると、図１５のように分布する。この平面上で、領域２および４に分布する非確定画素Ｐ１と領域１および３に分布する確定画素Ｐ２を図のような閾値Ｔ１を用いて分離する。この閾値Ｔ１より上の領域を非確定画素Ｐ１の領域、下の領域を確定画素Ｐ２の領域として分離処理を行う。
【００５５】
図４の一次微分−濃度値平面上での閾値Ｔ１の設定処理４０２では、入力である多値文字画像Ｉ１の最大と最小の濃度値の平均値を横方向の座標、縦方向は０座標の点を基準としてその基準点から直線的に区切った線が閾値Ｔ１となる。閾値Ｔ１の決定の際は、上記基準点から直線状に存在する画素の分布を探索し、最小となる２つの直線を求めてその直線を閾値Ｔ１とする。閾値Ｔ１を求めた後、閾値Ｔ１での分離処理４０３によって、入力である多値文字画像Ｉ１から、この一次微分−濃度値平面で閾値Ｔ１よりも上にある領域を非確定画素Ｐ１の領域として抽出する。そして、文字境界領域の分離画像Ｉ２を得る。
【００５６】
［第３の実施例］
図５に、第１の実施例で説明した文字境界領域の抽出処理１０１に対して、第３の実施例における画像処理方法での非確定画素の分離処理の流れを示す。この非確定画素Ｐ１の分離処理は、文字画像の濃度断面から領域を仮定し、スキャナ入力によって生じる文字境界領域に分布する非確定画素Ｐ１を抽出する。第２の実施例の図３に示した領域１〜４を、図１５に示す一次微分−濃度値平面上でＬＢＧアルゴリズムを用いて効率的に分離する。
【００５７】
第３の実施例での非確定画素の分離処理において、入力画像は第２の実施例と同じスキャナ入力の多値文字画像Ｉ１であり、最大と最小の一次微分値と、最大と最小の濃度値を抽出する。一次微分の強度は、第２の実施例の一次微分値の算出処理４０１と同様の処理を行って得る。
【００５８】
次に、一次微分−濃度値平面上でのＬＢＧアルゴリズムによるグルーピング処理５０２を行って、非確定画素Ｐ１を分離してゆく。図１５に示している領域を５つに分離するために、５領域の中心となるＬＢＧアルゴリズムの初期点を図１６のように設定する。次に、ＬＢＧアルゴリズムによって全分布点について、５点間との最小距離を算出してどの中心点に最も近いか距離計算し、最近点にグループ化してゆく。図１７に、それぞれの中心点と分布点の距離計算式を示す。一次微分−濃度値平面で濃度値をｘ方向、一次微分値をｙ方向として分布点の座標を（ｘ，ｙ）、それぞれの領域の中心点の座標を（ｘ₀，ｙ₀）として、図２１の下部計算式により分布点と中心点との距離計算を行う。グループ化された点でそのグループの中心点を算出し、再び最近点を見つけるため距離計算を行って最近点を探索して再びグループ化する。この中心点の算出演算を繰り返し行い、中心点の移動距離が収束したら処理を終了する。そのとき、領域１、３の中心点にグループ化されている画素を確定画素Ｐ２、領域２、４の中心点にグループ化されている画素を非確定画素Ｐ１として分離する。そして、文字境界領域の分離画像Ｉ２を得る。
【００５９】
［第４の実施例］
図６に、第１の実施例の構成で説明した文字境界領域の二値化処理１０２に対して、本発明の第４の実施例における画像処理方法での非確定画素の二値化処理の流れを示す。この非確定画素二値化処理は、文字境界領域の画素に対し、濃度断面の情報をもとに局所的な二値化を行う。
【００６０】
第４の実施例での二値化処理の入力は、スキャナ入力による多値文字画像Ｉ１と、文字境界領域の分離画像Ｉ２を用いる。
【００６１】
図１８に、文字画像の濃度断面の例を示す。スキャナ入力によって生じる文字境界領域に分布する非確定画素Ｐ１を対象にして、断面の勾配ごとに図のように閾値Ｔ２を設定し、その１濃度断面に属している非確定画素Ｐ１に対し、濃度断面という局所的な線状の情報を用いて二値化処理を行う。
【００６２】
図１９に、非確定画素Ｐ１の近傍領域を示す。文字画像は、２次元画像なので、４方向の濃度断面が存在する。これら４方向のスキャナ入力によって生じる文字境界領域の情報を用いて二値化処理を行う。まず、多値文字画像Ｉ１に対し、図６の方向微分オペレータによる一次微分の計算処理６０１を行う。画像に対し、横方向にラスタ走査を行い、注目する非確定画素Ｐを探索し、その画素Ｐを中心に、方向ごとに重みを付けたフィルタをかけて一次微分値（濃度断面の傾き）を計算する。これは、後で示すどの方向の二値化結果がより有効かを調べるための特徴量として計算する。濃度断面の傾きが大きい方向ほど、より文字境界領域で非確定画素Ｐがどこに分布しているかが明確で、二値化判定に有効だと考えられる。
【００６３】
一次微分算出処理６０１の後、＋４５°方向の非確定画素Ｐ１の近傍領域での二値化結果の取得処理６０２を行う。注目画素Ｐから確定画素Ｐ２までに存在する、その方向に分布する非確定画素Ｐ１を、近傍領域として、その近傍領域ごとに局所的な閾値Ｔ２を設定する。図２０に、図１９の＋４５°方向に分布する近傍領域と判断された非確定画素Ｐ１の濃度分布を示す。１方向ごとに近傍領域の閾値Ｔ２を設定し二値化処理する。
【００６４】
以後、この＋４５°方向の非確定画素Ｐ１の近傍領域での二値化結果の取得処理６０２と同様、方向を変えて同様の処理を行う。水平方向の非確定画素Ｐ１の近傍領域での二値化結果の取得処理６０３、−４５°方向の非確定画素Ｐ１の近傍領域での二値化結果の取得処理６０４、垂直方向の非確定画素Ｐ１の近傍領域での二値化結果の取得処理６０５を順に行い、４方向の方向別の二値化結果を得る。方向ごとの二値化結果に上記の一次微分値である濃度断面の傾きを投票処理６０６によって判定することによって、最終的な二値化結果を得る。全非確定画素Ｐ１と選択された画素に対し、上記処理を行う。この投票処理とは、例えば、水平方向の二値化結果が文字で一次微分値が５０、垂直方向の二値化結果が背景で一次微分値が１００、＋４５°方向の二値化結果が背景で一次微分値が１０、−４５°方向の二値化結果が文字で一次微分値が３５だとする。この場合、文字には水平方向と−４５°方向の一次微分値が投票され８５になり、背景には垂直方向と＋４５°方向の一次微分値が投票され１１０となり、この結果、最終的なこの画素の二値化結果は背景となる。
【００６５】
非確定画素Ｐ１と決定された全画素に対し、横方向のラスタ走査を行い探索し上記の処理を行い、非確定画素の領域のみの二値化画像Ｉ４を得る。
【００６６】
［第５の実施例］
図７に、第１の実施例で説明した文字境界領域以外の二値化処理１０３に対して、本発明の第５の実施例における画像処理方法での確定画素の二値化処理の流れを示す。この確定画素の二値化処理は、近傍の二値化された非確定画素Ｐ１の情報を用いて文字の内側の領域と背景領域に分布する確定画素Ｐ２を二値化処理する。
【００６７】
入力画像は、第１の実施例の非確定画素二値化処理２０２でスキャナ入力によって生じる文字境界領域が二値化されている非確定画素Ｐ１の領域のみの二値化画像Ｉ４を用いる。
【００６８】
確定画素Ｐ２の近傍領域の決定処理７０１によって確定画素Ｐ２の二値化に用いる領域を決定する。図２１に、確定画素Ｐ２の近傍領域を示す。このように注目している確定画素Ｐ２の８近傍画素を探索して、近傍領域として決定する。
【００６９】
近傍領域決定後、投票処理７０２を行って確定画素Ｐ２の二値化処理を行う。近傍領域と決定された場所に分布する二値化された非確定画素Ｐ１の個数を調べて投票し、多く投票された方に統合させて二値化する。また、数が同数の場合は、さらに、確定画素Ｐ２の近傍領域の拡大処理７０３を行い、近傍領域を図２１の右図のように拡大して１６近傍まで探索範囲を広げて同様に投票処理７０２を行い、二値化を行う。対象画像の左上から横方向にラスタ走査して確定画素Ｐ２を探索して上記の処理を行ってゆく。注目している確定画素Ｐ２の近傍領域に確定画素Ｐ２しか存在しない場合、既に処理された確定画素Ｐ２の二値化結果も投票に反映して二値化する。
【００７０】
図２２に、文字境界領域が二値化されており、確定画素Ｐ２は二値化されていない画像に対しての上記の処理の結果を示す。図２２のように背景に囲まれた確定画素Ｐ２は背景画素に、文字画素に囲まれた確定画素Ｐ２は文字画素に二値化される。
【００７１】
確定画素Ｐ１と判断された全ての画素に対し、上記の確定画素二値化処理を行い、二値文字画像（文字境界領域補正前）Ｉ５を得る。
【００７２】
［第６の実施例］
図８に、第１の実施例で説明した文字境界領域の補正処理１０４に対して、本発明の第６の実施例における画像処理方法での文字境界領域の補正処理の流れを示す。この文字境界領域の補正処理は、文字境界領域に生じる凹凸を平滑化によって埋める処理を行って除去する。
【００７３】
入力画像は、確定画素Ｐ２の二値化処理２０３後の全画素が二値化された二値文字画像（文字境界領域補正前）Ｉ５を用いる。
【００７４】
図１３に示したように、文字画像に対し、ｘ（横）方向に２回ラスタ走査して、横方向での文字画像の連続領域の間に分布する背景画素の穴埋め処理８０１と、横方向での背景画素の連続領域の間に分布する文字画素の穴埋め処理８０２とを行う。最初のｘ方向ラスタ走査で、文字画素を探索し、文字画素の連続領域を抽出したら、その方向に背景画素が続いているかどうか探索する。続いているならば、次の文字画素の連続領域をｘ方向の最後の画素まで探索する。続いていない場合、背景画素の後に再び文字画像が分布しているので、その背景画素を文字画素に埋めて文字画素の連続領域を繋ぐ処理を行う。全ｘ方向についてこの処理を行う。次に、背景画素に注目し、背景画素の連続領域に対して上記と同様の処理を行い、背景画素の連続領域を繋ぐ処理を全ｘ方向について行う。
【００７５】
ｙ（縦）方向についても同様に、２回ラスタ走査して、縦方向の文字画素の連続領域の間の穴埋め処理８０３と、縦方向の背景画素の連続領域の間に分布する文字画素の穴埋め処理８０４とを行う。
【００７６】
図２３に、文字境界領域の平滑化処理の例を示す。図２３の（ａ）や（ｃ）で文字の境界領域に存在するノイズと思われる境界領域の凹凸を平滑化することによって除去し、図２３の（ｂ）や（ｄ）のような文字の境界領域が補正された二値文字画像Ｉ３を出力する。二値化された文字画像の境界に着目し、低解像度の影響により二値化処理の際に生じる境界領域の凹凸を抽出し、平滑化して除去する。縦方向の文字画素の連続領域を検出し、その連続領域の間隔が閾値以下の場合、その画素を文字画素に埋めることによって文字境界領域の凹凸を埋める。同様の処理を横方向の文字画素の連続領域に対しても行い、その連続領域の間隔が閾値以下の場合、その画素を文字画素に埋めることによって文字境界領域の凹凸を埋める。また、背景画素の縦と横方向の連続領域に対しても同様の処理を行い、文字画像の境界領域の凹凸を埋める。図２４に上記平滑化処理の実行例を示す。
【００７７】
［第７の実施例］
図９に、本発明の第７の実施例における画像処理方法での領域ごとに二値化処理を切り替える処理を示す。非確定画素Ｐ１の分離処理では、文字画像の濃度断面から領域を仮定し、スキャナ入力によって生じる文字境界領域に分布する非確定画素Ｐ１を抽出する。まず、セレクト処理９０１で、一次微分や濃度値などの特徴量を用いて、領域を分離する。そして、領域ごとに異なる二値化処理９０２、９０３を施す。
【００７８】
図１２に、第２の実施例の非確定画素の分離処理２０１を施した画像の濃度分布図を示す。一次微分と濃度情報とからこのように文字境界領域である非確定画素Ｐ１を抽出し、このスキャナ入力の影響を多く受けた画素に対し、第７の実施例の方法における局所的な閾値Ｔ２の設定を行い、二値化処理する。
【００７９】
また、図２５に、濃度断面での非確定画素Ｐ１と確定画素Ｐ２の分布図を示す。この図から、非確定画素Ｐ１は、文字境界領域に相当する文字境界領域に分布し確定画素Ｐ２は濃度変化が平坦な領域に分布することがわかる。スキャナ入力によって生じる文字境界領域に分布する画素と文字境界領域以外に分布する画素とを分離することによって、スキャナ入力によって生じる文字境界領域に分布している非確定画素Ｐ１に対して１濃度断面ごとに閾値Ｔ２を設定し、二値化処理する。
【００８０】
それぞれの二値化手法で二値化処理した後、その結果の統合処理９０４によって、セレクト処理９０１において分離され二値化処理された結果を統合して、その結果を最終的な二値文字画像Ｉ３として出力する。
【００８１】
［第８の実施例］
図１０に、第１の実施例で説明した文字境界領域の二値化処理１０２に対して、本発明の第８の実施例における画像処理方法での非確定画素の二値化処理の流れを示す。この非確定画素の二値化処理では、文字境界領域の画素に対し、濃度断面の情報をもとに局所的な二値化処理を行う。この処理の入力は、スキャナ入力による多値文字画像Ｉ１と、文字境界領域の分離画像Ｉ２とを用いる。前述の第４の実施例とは、＋４５°方向の非確定画素Ｐ１の近傍領域での新しい閾値による二値化結果取得処理１００２から垂直方向の非確定画素Ｐ１の近傍領域での新しい閾値による二値化結果取得処理１００５までの処理における閾値設定方法が異なる。
【００８２】
＋４５°方向の非確定画素の近傍領域での新しい閾値による二値化結果取得処理１００２では、近傍領域を、確定画素Ｐ２の位置までに存在する、その方向に分布する非確定画素Ｐ１を、近傍領域として、その近傍領域ごとに、局所的な閾値を設定する。図２６に、文字の掠れや潰れを防ぐための閾値シフトを示す。第４の実施例では、近傍領域として決定した非確定画素Ｐ１の濃度値を領域し、最大と最小濃度の平均値を閾値Ｔ２に設定しいてる。これに対し、第８の実施例では、注目画素が分布する非確定画素Ｐ１の濃度分布の形状を利用して閾値Ｔ２をシフトして、濃度断面の小さな山や谷を再現できるように設定し、文字の掠れや潰れを防ぐ。図２６の（ａ）のように、濃度断面の形状が山を描くように分布している場合、文字の掠れを防ぐために閾値を下げる。（ｂ）のように、濃度断面の形状が谷を描くように分布している場合、文字の潰れを防ぐために閾値を上げる。それ以外の形状、単調増加や減少の場合は、閾値のシフトは行わない。
【００８３】
投票処理１００６では、第４の実施例と同様な処理を行い、最終的な二値化結果を得る。画像の横方向に非確定画素Ｐ１を探索し、全非確定画素Ｐ１に対して上記の処理を行い、非確定画素Ｐ１の領域のみの二値化画像Ｉ４を得る。
【００８４】
以上により本発明の実施形態について説明した。なお、上述した実施形態は、本発明の好適な実施形態の一例を示すものであり、本発明はそれに限定されるものではなく、その要旨を逸脱しない範囲内において、種々変形実施が可能である。
【００８５】
【発明の効果】
以上の説明から明らかなように、請求項１記載の発明によれば、従来の二値化手法では二値化判定を誤る確率の高いスキャナ入力によって生じる文字境界領域を抽出し、二値化対象画素が文字画像のどこに分布しているかを濃度断面の情報から判断して二値化しているので、出力後の文字画像の画質を向上することができる。また、また、ＯＣＲ認識率を向上させることができる。
【００８６】
また、請求項２記載の発明によれば、文字画像からスキャナ入力によって生じる文字境界領域を分離する。従来の二値化手法では二値化判定を誤る確率の高いスキャナ入力によって生じる文字境界領域に対して二値化処理を換えることが可能となる。
【００８７】
また、請求項３記載の発明によれば、文字画像からスキャナ入力によって生じる文字境界領域を、ＬＢＧアルゴリズムを用いることによって効率良く分離する。ＬＢＧアルゴリズムは最初に設定した初期点の個数だけ領域を分割することが可能である。従来の二値化手法では二値化判定を誤る確率の高いスキャナ入力によって生じる文字境界領域に対して二値化処理を換えることが可能となる。
【００８８】
また、請求項４記載の発明によれば、スキャナ入力によって生じる文字境界領域に対し、１濃度断面の情報を用いて局所的な閾値設定が可能となる。１濃度断面から閾値を設定するので、太い文字線に囲まれた背景領域を抽出したり、細い文字線を忠実に再現することが可能となる。また、ＯＣＲ認識率が向上する。
【００８９】
また、請求項５記載の発明によれば、スキャナ入力によって生じる文字境界領域以外の領域において、文字の内側に背景と誤判定したり、背景領域に文字領域が存在しにくくなるので、出力後の文字画像の画質が向上するとともに、文字領域内の粒状の文字画像のノイズに敏感なＯＣＲにおいて認識率が向上する。
【００９０】
また、請求項６記載の発明によれば、二値化処理後に生じる文字境界領域の凹凸に対し、穴埋め処理による平滑化により除去を行うので、文字画像の画質が向上するとともに、文字境界領域の凹凸に敏感なＯＣＲにおいて認識率が向上する。
【００９１】
また、請求項７記載の発明によれば、スキャナ入力によって生じる文字境界領域ごとに、１濃度断面の情報を用いて局所的な閾値を変化させた局所的な二値化処理が可能となる。１濃度断面から閾値を設定するので、太い文字線に囲まれた背景領域を抽出したり、細い文字線を忠実に再現することが可能となる。また、ＯＣＲ認識率が向上する。
【００９２】
また、請求項８記載の発明によれば、スキャナ入力によって生じる文字境界領域に対し、文字の濃度断面とその形状の情報を用いた局所的な閾値設定が可能となる。文字濃度断面とその形状の情報から閾値を変えて設定するので、太い文字線に囲まれた背景領域を抽出したり、細い文字線を忠実に再現することが可能となる。また、ＯＣＲ認識率が向上する。
【図面の簡単な説明】
【図１】本発明の第１の実施の形態における画像処理方法を示す図である。
【図２】本発明の第１の実施の形態における画像処理方法の二値化処理の流れを示す図である。
【図３】本発明の第１の実施の形態における画像処理方法の二値化処理の別方法を示す図である。
【図４】本発明の第２の実施の形態における画像処理方法の非確定画素分離処理を示す図である。
【図５】本発明の第３の実施の形態における画像処理方法の非確定画素分離処理を示す図である。
【図６】本発明の第４の実施の形態における画像処理方法の二値化処理を示す図である。
【図７】本発明の第５の実施の形態における画像処理方法の確定画素の二値化処理を示す図である。
【図８】本発明の第６の実施の形態における画像処理方法の文字境界領域の補正処理を示す図である。
【図９】本発明の第７の実施の形態における画像処理方法の二値化処理切り替え処理を示す図である。
【図１０】本発明の第８の実施の形態における画像処理方法の非確定画素二値化処理を示す図である。
【図１１】単純閾値Ａと局所的閾値Ｂによるそれぞれの二値化処理の例を示す図である。
【図１２】非確定画素の分離処理を施した画像を示す図である。
【図１３】文字境界領域の凹凸の除去を示す図である。
【図１４】文字画像の濃度断面について示す図である。
【図１５】一次微分と濃度値の分布を示す図である。
【図１６】ＬＢＧアルゴリズムによる非確定画素Ｐ１の分離処理を示す図である。
【図１７】中心点と分布点との距離計算式について示す図である。
【図１８】文字画像の濃度断面の例を示す図である。
【図１９】非確定画素Ｐ１の近傍領域を示す図である。
【図２０】＋４５°方向に分布する近傍領域と判断された非確定画素Ｐ１の濃度分布を示す図である。
【図２１】確定画素Ｐ２の近傍領域を示す図である。
【図２２】確定画素Ｐ２の二値化処理の実行例を示す図である。
【図２３】文字境界領域の平滑化処理についての図である。
【図２４】文字境界領域の平滑化処理の実行例を示す図である。
【図２５】濃度断面での非確定画素Ｐ１と確定画素Ｐ２の分布について示す図である。
【図２６】文字の掠れや潰れを防ぐための閾値のシフトを示す図である。
【図２７】一般的な文字読み取り装置の構成を示す図である。
【図２８】解像度の違いによる二値化処理後の文字境界領域の凹凸を示す図である。
【符号の説明】
Ｉ１スキャナ入力の多値文字画像
Ｉ２文字境界領域の分離画像
Ｉ３二値文字画像
Ｉ４非確定画素の領域のみの二値化画像
Ｉ５二値文字画像（文字境界領域補正前）
Ｐ１非確定画素
Ｐ２確定画素
Ｔ１非確定画素の分離処理に用いる閾値
Ｔ２非確定画素の近傍領域の閾値
１スキャナ部
２画像記憶部
３ＣＰＵ（画像処理部）
４ＯＣＲ（文字認識処理部）[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an image processing apparatus and method, and more particularly to an image processing apparatus and method for performing character image binarization processing for character recognition processing.
[0002]
[Prior art]
As digitization of documents has progressed, character information formed on paper has been digitized and stored as image information. However, image information in a format such as a bitmap has a large amount of data, and further, it is difficult to edit a document (image information). Therefore, conversion to a character code is desired.
[0003]
Therefore, an OCR (Optical Character Reader: optical character reader) that reads a character image and converts it into a character code is used. When performing processing to convert a character code by OCR, the character image is cut out for each character image. Since the character image is normalized and recognized based on the direction of the edge of the character boundary region, the position of the character pixel is first identified in the multivalued character image digitized by an input device such as a scanner. Therefore, it is necessary to separate the character pixels and the background pixels.
[0004]
It is necessary to perform a binarization process using a character image multi-valued by scanner input as an input and output a character image expressed in a binary form. FIG. 27 shows the configuration of a device that performs a general character image binarization process. This apparatus digitizes character information formed on a paper surface and reads it as image data of multi-value expression, an image for storing multi-value image data read by the scanner unit 1 and image data after image processing. A storage unit 2, a CPU (signal processing unit) 3 that performs binarization processing on the input multi-value image data, and an OCR 4 that performs character recognition processing based on the character image data expressed in binary form. Also, the arrows in FIG. 27 indicate the flow of image information.
[0005]
The flow of processing is as follows. The multi-value character image read by the scanner unit 1 is temporarily stored in the image storage unit 2. Next, the CPU 3 performs processing for separating the character boundary region generated by the scanner input on the multi-value character image, and stores the separated image of the character boundary region in the image storage unit 2. Based on the input multi-value character image stored in the image storage unit 2 and the separated image of the character boundary area created by image processing in the CPU 3, the input multi-value character image in the CPU 3 A valuation process and a smoothing process of the character boundary area are performed. As a result of the binarization process and the character boundary area smoothing process, the binary character image data obtained is stored in the image storage unit 2 again. After storing the binary character image in the image storage unit 2, the OCR 4 performs character recognition processing using the binary character image as an input. In the OCR4, binarization is performed with a simple threshold value. Therefore, an 8-bit image in which a pixel determined to be a character area has a density value of 255 and a pixel determined to be a background pixel has a density value of 0 is used for character recognition. The image is sent to the OCR 4 as a target image.
[0006]
As a conventional technique for binarizing a character or a document image, Japanese Patent Laid-Open No. 2000-333022 discloses an image binarization method and apparatus, and a storage medium. Japanese Patent Laid-Open No. 5-282494 discloses a binarization apparatus for image data. Japanese Patent Laid-Open No. 10-308771 discloses a binarization method and a character reading device. These prior arts are binarization methods and apparatuses for text and document images, but use density information as information when performing binarization processing.
[0007]
For this reason, character images read by an input device, particularly a scanner, are converted to multi-values by means of CCD reading or MTF characteristics, even if the original binary character information is due to scanner characteristics, particularly in character boundary regions where the density gradient is steep. Concentration profile is rounded. Since the scanner characteristics are unique to each scanner device, and the degree of rounding of the density cross section differs for each scanner, the gradient of the density value edge changes in a multivalued character image. In particular, an intermediate density is likely to occur at a density cross section of a multi-valued image input by a scanner having a sharpening degree. During binarization processing, it is difficult to determine whether to binarize the pixels distributed in the intermediate density area. With the conventional binarization method, the character image may be crushed or drowned due to incorrect binarization determination of the boundary area. There is a problem of doing.
[0008]
In addition, the degree of rounding of the density cross section is high because the local area information, for example, the area surrounded by a thick character line, is a background area, and the density value is affected by the nearby character area. There is a tendency to grow. The area resulting from the darkening of the density cross section due to the scanner input is mainly the character boundary area, which is a transition area from the character pixel to the background pixel or from the background pixel to the character pixel. To do.
[0009]
In particular, sampling becomes rough due to the influence of low resolution, and when digitization is performed, variations in density values are likely to occur particularly in character boundary regions where density changes are large. FIG. 28 shows how unevenness occurs in the character boundary area in the binarized image due to the difference in resolution of the multilevel image in the binarization processing of the multilevel image with different resolutions. The left side shows an image captured at high resolution by the scanner. The right side shows an image captured by the scanner at a low resolution. In particular, in the lower-resolution image in the upper right diagram, since the sampling interval is large, a steep density gradient in the character boundary region cannot be expressed, and density variation occurs. The results of binarizing these images with simple threshold values (density value 128, * 1 pixel 8-bit expression) are shown in the lower diagram. As described above, a sharp density gradient in the character boundary region cannot be expressed particularly in a low-resolution image. Therefore, unevenness is likely to occur in an image after binarization by the conventional binarization method.
[0010]
When binarization is performed using only density information for a multi-value character image input by a scanner, as described above, the degree of rounding of the density cross-section caused by the scanner characteristics is locally different. There is a problem that the OCR recognition rate is reduced due to collapse or drowning. In the conventional binarization method, the density cross section generated by the scanner input is distorted, and the binarization determination is often wrong in the region resulting from the density cross section being distorted. In order to improve the OCR recognition rate, there is a need for a binarization processing method and apparatus that makes it difficult to make an erroneous determination with respect to a region that is generated as a result of a loss of density cross section.
[0011]
In addition, as described above, there is a problem that unevenness tends to appear in the character boundary area due to low resolution input, and the OCR recognition rate is lowered. OCR performs recognition by detecting the direction of the boundary of the character in the boundary region of the character image. A method and apparatus capable of faithfully binarizing information in a character boundary region, in which a characteristic amount for recognition is used in order to improve an OCR recognition rate, and there are many misjudgments in binarization processing in the conventional method. Is required.
[0012]
[Problems to be solved by the invention]
The present invention has been made in view of such a problem. For character recognition processing in OCR, the character boundary region is faithful as much as possible in binarization processing for outputting a binary character image from a multi-value character image. It is an object of the present invention to provide an image processing method and apparatus that can be binarized to improve the OCR recognition rate.
[0013]
According to the first aspect of the present invention, the binarization misjudgment of the character boundary region is performed by using the density cross-section information that is local information of the character for the binarization determination with respect to the character boundary region generated by the scanner input. The purpose is to suppress and improve the OCR recognition rate.
[0014]
Since the character image has a high contrast, the character image input by the scanner is particularly liable to be blurred due to the deterioration of the character boundary region. The boundary area of characters generated by scanner input is particularly affected by the local position information of the character line and the thickness of the font because the degree of deterioration of the boundary area of the character changes depending on the density information of the neighboring area. In the conventional method in which the threshold value of the density value is set to a single pixel, the characters are crushed or blurred after binarization. According to the second aspect of the present invention, in order to prevent an erroneous determination of binarization of a character boundary region, a character boundary region generated by scanner input is extracted, binarization suitable for this region is performed, and an OCR recognition rate is obtained. It aims at improving.
[0015]
Since the character image has a high contrast, the character image input by the scanner is particularly liable to be blurred due to the deterioration of the character boundary region. Since the boundary area of the character generated by the scanner input also changes depending on the local position information of the character line and the thickness of the font, especially in the conventional method in which a single density value threshold is set for all the pixels. Characters were crushed and drowned after valuation. According to the third aspect of the present invention, by using the LBG algorithm for pixels that are difficult to determine for binarization, the character boundary region generated by this scanner input is extracted efficiently, and is suitable for this region. An object is to perform binarization processing and improve the recognition rate of OCR.
[0016]
In the character boundary region generated by the scanner input, character pixels and background pixels are mixed. In the invention according to claim 4, in order to binarize the area, a local binarization threshold is provided for each density cross section, the binarization process is performed for each density cross section, and local binarization is performed. An object of the present invention is to prevent the character image after binarization from being crushed or blurred by setting the binarization threshold and to improve the OCR recognition rate.
[0017]
The invention according to claim 5 stably binarizes an area inside a character and a background area using an adjacent binarization result for an image in which a character boundary area generated by scanner input is binarized. The purpose is to process.
[0018]
When it is desired to perform character recognition by OCR, it may be erroneously recognized due to unevenness generated in the character boundary region after binarization processing, particularly due to the effect of low resolution. An object of the present invention is to improve the recognition rate of OCR by removing unevenness generated in the boundary region by smoothing the character boundary region from the binarized character image.
[0019]
The character boundary region generated by the scanner input tends to be deteriorated due to the deterioration of the character boundary region due to the influence of the scanner characteristics and the character font. Since the character image read by a scanner or the like is also affected by the character vicinity region, the density value differs greatly between the character region and the background region. If an image input by a scanner is binarized by setting a single density threshold value for all pixels, characters may be crushed or blurred. According to the seventh aspect of the invention, the binarization error in the binarized character boundary electronic region is obtained by separating the character boundary region caused by the scanner input and switching the binarization processing for each region by the detection means. The purpose is to suppress the determination and improve the recognition rate of OCR.
[0020]
In the character boundary region generated by the scanner input, character pixels and background pixels are mixed. In the invention described in claim 8, in order to binarize the region, a local binarization threshold is provided for each cross section, and binarization processing is performed for each density cross section. An object of the present invention is to prevent crushing and blurring of a character image after binarization and improve the recognition rate of OCR by setting a local binarization threshold using the shape of the density cross section.
[0021]
[Means for Solving the Problems]
In order to achieve such an object, the invention described in claim 1 includes a scanner unit that digitizes character information formed on a paper surface and captures it as multi-value image information, an arithmetic processing unit that performs an image processing calculation, An image processing apparatus comprising: an information storage means for storing information and an image processing result obtained by an arithmetic processing means, a separating means for separating a character boundary area generated by an input by a scanner means; and binarizing the character boundary area First binarizing means for processing, second binarizing means for binarizing an area other than the character boundary area, and correcting means for smoothing and correcting irregularities in the character boundary area after binarization processing An image processing apparatus comprising: The first binarization means is a character boundary region With respect to horizontal, vertical, ± 45 degrees direction Each non-deterministic pixel distributed from the target pixel to the definitive pixel is set as a neighboring area, and binarization processing is performed on the character and the background based on a threshold value set according to the average density value of the neighboring area. ± 45 degree direction of Above Neighboring region for binarization result of Density value and density gradient of text image Is the first derivative based on Weighting the extraction results Compare the sum of the first derivative values in all directions where the binarization result is text with the sum of the first derivative values in all directions where the binarization result is the background. Final binarization result As character or as background It is characterized by a voting process to decide.
[0022]
The invention described in claim 2 is characterized in that, in the invention described in claim 1, the separating means separates the character boundary region using the information of the primary differential value and the density value as the feature amount.
[0023]
The invention described in claim 3 is characterized in that, in the invention described in claim 1, the separating means separates the character boundary region by the LBG algorithm using the first-order differential-density plane of the input image.
[0024]
According to a fourth aspect of the present invention, in the first aspect of the invention, the first binarizing unit binarizes the character boundary region using the character density section information generated by the input by the scanner unit. It is a feature.
[0025]
The invention according to claim 5 is characterized in that, in the invention according to claim 1, the second binarization means performs binarization processing using the binarization processing result of the neighboring character boundary region. .
[0026]
The invention described in claim 6 is characterized in that, in the invention described in claim 1, the correction means performs smoothing by detecting and filling in the unevenness generated in the character boundary region of the binarized image. .
[0027]
The invention described in claim 7 is characterized in that, in the invention described in claim 1, character boundary regions are separated and binarization processing is switched for each separated region.
[0028]
According to an eighth aspect of the present invention, in the first aspect of the invention, the first binarizing unit binarizes the character boundary area by using the character density cross section generated by the input by the scanner unit and the information on the shape thereof. It is characterized by processing.
[0029]
According to a ninth aspect of the present invention, a separation step of separating a character boundary region generated by scanner input and a first binarization step of binarizing the separated character boundary region for a scanner input multi-value character image And a second binarization step for binarizing an area other than the character boundary area, and a correction step for smoothing and correcting irregularities in the character boundary area after the binarization process. , In the first binarization step, the character boundary region With respect to horizontal, vertical, ± 45 degrees direction Each non-deterministic pixel distributed from the target pixel to the definitive pixel is set as a neighboring area, and binarization processing is performed on the character and the background based on a threshold value set according to the average density value of the neighboring area. ± 45 degree direction For the binarization result of of Density value and density gradient of text image Is the first derivative based on Weighting the extraction results Compare the sum of the first derivative values in all directions where the binarization result is text with the sum of the first derivative values in all directions where the binarization result is the background. Final binarization result As character or as background It is characterized by a voting process to decide.
[0030]
The invention described in claim 10 is characterized in that, in the invention described in claim 9, the separation step separates the character boundary area using the information of the primary differential value and the density value as the feature amount.
[0031]
The invention described in claim 11 is characterized in that, in the invention described in claim 9, the separation step separates the character boundary region by the LBG algorithm using the first-order differential-density plane of the input image.
[0032]
The invention described in claim 12 is characterized in that, in the invention described in claim 9, the first binarization step binarizes the character boundary region using the character density section information generated by the scanner input. Yes.
[0033]
The invention described in claim 13 is characterized in that, in the invention described in claim 9, the second binarization step performs binarization processing using the binarization processing result of the neighboring character boundary region. .
[0034]
The invention described in claim 14 is characterized in that, in the invention described in claim 9, the correction step performs correction by detecting and filling in the unevenness generated in the character boundary region of the binarized image.
[0035]
The invention described in claim 15 is characterized in that, in the invention described in claim 9, character boundary regions are separated and binarization processing is switched for each separated region.
[0036]
According to a sixteenth aspect of the present invention, in the ninth aspect of the invention, the first binarization step binarizes the character boundary region by using the character density cross section generated by the scanner input and information on the shape thereof. It is characterized by doing.
[0037]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.
[0038]
[First embodiment]
FIG. 1 is a diagram showing an image processing method according to the first embodiment of the present invention. This image processing is processing performed by a CPU (signal processing unit). In this image processing method, the character boundary region extraction process 101 for the input multi-value character image I1 and the character boundary region separation image I2 obtained after the multi-value character image I1 and the character boundary region extraction processing are sequentially performed. Character boundary region binarization processing 102 based on the above, binarization processing 103 for regions other than the character boundary region for the image after the character boundary region binarization processing, and output after the above binarization processing A character boundary region correction process 104 is performed on the image, and a binary character image I3 is output.
[0039]
In the character boundary region extraction processing 101, a character boundary region generated by scanner input is separated based on a feature amount such as a first derivative (of a density value curve).
[0040]
In the binarization processing 102 for the character boundary region, attention is paid to the character boundary region generated by the scanner input, and the binarization processing is performed by locally changing the determination threshold. FIG. Shows an example of Mincho and Gothic character images captured from the scanner and an example of the binarization result. FIG. The upper image shows the original image. The upper and lower figures in the center show the density distribution on the density section of the original image. The lower image shows an image obtained as a result of binarization processing using the threshold A and the threshold B shown on the density distribution. In this process, FIG. As described above, raster scanning is performed in the horizontal direction of the target character image, and a density cross section is set for each horizontal direction. The edge portion which is a character boundary region uses an image obtained as a result of separation by the character boundary region extraction processing 101 described above.
[0041]
The threshold A in the center diagram is a simple threshold set to 128, which is an intermediate density value of an 8-bit image. On the other hand, the threshold value B is set to an intermediate value between the maximum density pixel and the minimum density pixel in the area (dotted line part in the upper part of the center diagram) determined as the edge part that is the character boundary area in the character boundary area extraction processing 101. It is set. Similarly, the threshold value B is set for the edge portion which is another character boundary region. FIG. The lower image is a binarization processing result after the above processing is fully scanned in the horizontal direction. FIG. As can be seen by referring to the lower image, when binarization is performed by setting a simple threshold A from the entire image, the character border area generated by the scanner input is drowned in the Mincho style with thin character lines, The Gothic font with thick character lines is crushed. On the other hand, in the method in which the local threshold value B is set for each density section (edge portion), the peaks and valleys of the entire density section can be reproduced.
[0042]
Actually, in this binarization processing 102, only the edge portion which is the character boundary region is binarized. However, for comparison with the conventional method based on the simple threshold, pixels other than the character boundary region are threshold A. An example of binarization is shown.
[0043]
In the binarization processing 103 of the area other than the character boundary area, the character boundary area generated by the scanner input in the above process 102 is binarized. By using and integrating the information of the character boundary region generated by the digitized scanner input, the binarization process is stably performed.
[0044]
In the character boundary region correction process 104, sampling is particularly rough due to the low resolution, and the character boundary region where the density value is likely to vary in the character boundary region where the density change is particularly large when digitized. Perform correction processing. Since the low-resolution image cannot express the steep density gradient of the character boundary area, unevenness is likely to occur in the binarized image. It is removed by smoothing process.
[0045]
FIG. 2 is a diagram showing the flow of binarization processing based on the basic configuration of the image processing method of the present invention shown in FIG. In contrast to the character boundary region extraction processing 101 in FIG. 1, in the processing in FIG. 2, non-deterministic pixel separation processing 201 is performed. As an input image, a multi-value character image captured by the scanner unit has a sudden density change in the character boundary region, and pixels distributed in the character boundary region caused by the scanner input are set as non-deterministic pixels P1, and the character boundary region In other cases, the separation process is performed with the pixels in which the density change in the vicinity is distributed in a flat region as the determined pixel P2.
[0046]
FIG. 12 shows a diagram of an image subjected to the non-deterministic pixel separation process 201. The figure on the left is an example of a multi-value character image captured by scanner input, and an image obtained by subjecting the original image to the non-deterministic pixel separation process 201 is a figure on the right. The non-deterministic pixel P1 is shown in white and the definite pixel P2 is shown in black. As shown in the right figure, the non-deterministic pixels P1 are character boundary regions generated by the scanner input, and are therefore distributed in the boundary region where characters are blurred due to the scanner characteristics. Since the definite pixels P2 are regions other than the character boundary region generated by the scanner input, they are distributed in the region inside the character line and the background region that is not affected by the characters. Pixels distributed mainly in the character boundary area are defined as non-deterministic pixels P1. The pixels distributed in the background area not affected by the characters and the area inside the character line are defined as the confirmed pixels P2. As a specific separation method, the non-deterministic pixel P1 uses the fact that the density change in the vicinity is abrupt, and performs the separation process using the first derivative characteristic of the density as a feature amount. An edge detection filter is applied to the input multi-value character image to calculate the intensity of the first derivative. A threshold is provided for the intensity of this first derivative, and a separated image 208 of the character boundary region is obtained with pixels above the threshold as non-determinable pixels P1 and pixels below the threshold as definite pixels P2.
[0047]
Next, the binarization process 202 of the non-deterministic pixel P1 is performed as the binarization process 102 of the character boundary region. The input is a multi-value character image captured by the scanner unit and a separated image of the character boundary region. The multi-value character image is used for determining a threshold value from density information. The separated image of the character boundary region is used to check whether the target image is the non-deterministic pixel P1 distributed in the character boundary region. A local square neighborhood is set around the indeterminate pixel of interest, and binarization is performed using the average density value of the pixels distributed in the neighborhood as the threshold of the indeterminate pixel of interest. Process.
[0048]
Next, as a binarization process 103 for an area other than the character boundary area, a binarization process 203 for the definite pixel P2 is performed. As an input, an image obtained by binarizing only the character boundary region generated by the scanner input after the binarization processing 202 of the non-deterministic pixel P1 is used. In order to binarize the peripheral region of the character boundary region in which the non-deterministic pixel P1 is distributed, a character is displayed in the vicinity of the definite pixel P2 based on the information of the non-deterministic pixel P1 distributed in the vicinity of the definite pixel of interest. It is determined by calculating whether there are many pixels or many background pixels, and integration is performed on the larger one.
[0049]
Next, the character boundary region correction processing 204 is performed on the character boundary region correction processing 104 of the image processing method in the first embodiment. As an input, an image obtained by binarizing all pixels in the binarization process 203 of the definite pixel P2 is used. When raster scanning is performed from the upper left pixel in the X direction (horizontal) and Y direction (vertical) as shown in FIG. 13 to detect a continuous area of character pixels in the horizontal and vertical directions, and the gap between the continuous areas is narrow Then, the unevenness of the character boundary in the horizontal and vertical directions is filled by filling the pixel into the character pixel. In FIG. 13, in the examples 2 and 4, white pixels (background pixels) generated during the continuation of black pixels (character pixels) are corrected to black pixels (character pixels).
[0050]
Next, a continuous area of background pixels in the horizontal and vertical directions is detected from the upper left pixels in the X and Y directions as shown in FIG. 13, and if the space between the continuous areas is narrow, the pixels are filled in the background pixels. Thus, the unevenness of the character boundary in the horizontal and vertical directions is filled. In FIG. 13, in the examples 1 and 3, the black pixel (character pixel) generated during the continuation of the white pixel (background pixel) is corrected to the white pixel (background pixel). A binary character image in which the character boundary region is corrected by filling the unevenness of the boundary region of the character image is output.
[0051]
Also, FIG. 3 shows a flow of binarization processing of a variation different from the image processing method of FIG. The difference from the method of FIG. 2 is that in the binarization process 303 of the definite pixel P2, the definite pixel P2 is obtained by using the input multi-value character image I1 and the binarized image of only the non-deterministic pixel P1. The process of binarizing is performed. Since the definite pixel P2 can be separated into whether the density value is sufficiently high or low, the definite pixel P2 is binarized by providing a threshold value for the density value without using neighboring information as in the method of FIG. The other processes are the same as the method of FIG.
[0052]
[Second Embodiment]
FIG. 4 shows separation processing of the non-deterministic pixel P1 in the image processing method according to the second embodiment of the present invention, in contrast to the character boundary region extraction processing 101 described in the first embodiment. In this non-deterministic pixel P1 separation process, an area is assumed from the density cross-section of the character image, and the multi-value character image I1 is input, and non-deterministic pixels P1 distributed in the character boundary region generated by the scanner input are extracted.
[0053]
FIG. 14 shows a density cross section of a character image. The region 1 is a background region with a flat density gradient, the region 2 is a character-to-background / background-to-character transition region, the region 3 is a character region with a flat density gradient, and the region 4 is a character line. And a transition area to a narrow background area sandwiched between character lines. First, it is assumed that there is a region as shown in FIG. 14, and regions 2 and 4 in which non-deterministic pixels P1 are distributed are extracted based on the intensity and density information of the first derivative. The strength of the first derivative is obtained by performing a first derivative amount calculation process 401.
[0054]
FIG. 15 shows the distribution of the first derivative and the density value. Each region 1 to 4 in FIG. 14 is distributed as shown in FIG. 15 when viewed on the first-order differential-concentration value plane. On this plane, the non-deterministic pixels P1 distributed in the regions 2 and 4 and the definite pixels P2 distributed in the regions 1 and 3 are separated using a threshold value T1 as shown in the figure. The separation process is performed with the region above the threshold T1 as the region of the non-deterministic pixel P1 and the region below the region of the definite pixel P2.
[0055]
In the setting process 402 of the threshold value T1 on the primary differential-density value plane in FIG. 4, the average value of the maximum and minimum density values of the input multi-value character image I1 is the horizontal coordinate, and the vertical direction is the zero coordinate. A line that is linearly separated from the reference point with the point as a reference is the threshold value T1. When the threshold value T1 is determined, the distribution of pixels existing in a straight line from the reference point is searched, two minimum straight lines are obtained, and these straight lines are set as the threshold value T1. After obtaining the threshold value T1, the region above the threshold value T1 in the primary differential-density value plane is determined as the region of the non-deterministic pixel P1 from the input multi-value character image I1 by the separation process 403 at the threshold value T1. Extract. Then, a separated image I2 of the character boundary region is obtained.
[0056]
[Third embodiment]
FIG. 5 shows the flow of non-deterministic pixel separation processing in the image processing method according to the third embodiment, in contrast to the character boundary region extraction processing 101 described in the first embodiment. In this non-deterministic pixel P1 separation process, a region is assumed from the density cross section of the character image, and non-deterministic pixels P1 distributed in the character boundary region generated by the scanner input are extracted. The regions 1 to 4 shown in FIG. 3 of the second embodiment are efficiently separated using the LBG algorithm on the primary differential-concentration value plane shown in FIG.
[0057]
In the non-deterministic pixel separation process in the third embodiment, the input image is a multi-value character image I1 of the same scanner input as in the second embodiment, and the maximum and minimum primary differential values and the maximum and minimum densities. Extract the value. The intensity of the first derivative is obtained by performing the same process as the first derivative calculation process 401 in the second embodiment.
[0058]
Next, the nondeterministic pixel P1 is separated by performing grouping processing 502 by the LBG algorithm on the primary differential-density value plane. In order to separate the area shown in FIG. 15 into five areas, the initial point of the LBG algorithm that is the center of the five areas is set as shown in FIG. Next, the minimum distance between the five points is calculated for all distribution points by the LBG algorithm, the distance closest to the center point is calculated, and the points are grouped to the nearest point. FIG. 17 shows a distance calculation formula between the center point and the distribution point. On the primary differential-concentration value plane, the density value is in the x direction, the primary differential value is in the y direction, the coordinates of the distribution point are (x, y), and the coordinates of the center point of each region are (x ₀ , Y ₀ ), The distance between the distribution point and the center point is calculated by the lower calculation formula of FIG. The center point of the group is calculated from the grouped points, and in order to find the nearest point again, distance calculation is performed to search for the nearest point and group again. This calculation process of the center point is repeatedly performed, and the process ends when the movement distance of the center point converges. At that time, the pixels grouped at the central points of the regions 1 and 3 are separated as the definite pixels P2, and the pixels grouped at the central points of the regions 2 and 4 are separated as the non-deterministic pixels P1. Then, a separated image I2 of the character boundary region is obtained.
[0059]
[Fourth embodiment]
FIG. 6 shows the binarization processing of the non-deterministic pixel in the image processing method according to the fourth embodiment of the present invention, compared to the binarization processing 102 of the character boundary area described in the configuration of the first embodiment. Show the flow. In this non-deterministic pixel binarization process, local binarization is performed on the pixels in the character boundary region based on the density cross-section information.
[0060]
In the binarization processing in the fourth embodiment, a multi-value character image I1 obtained by scanner input and a separated image I2 of the character boundary region are used.
[0061]
FIG. 18 shows an example of a density cross section of a character image. For the non-deterministic pixel P1 distributed in the character boundary region generated by the scanner input, a threshold value T2 is set for each cross-sectional gradient as shown in the figure, and the density is determined for the non-deterministic pixel P1 belonging to the one density cross-section. Binarization processing is performed using local linear information called a cross section.
[0062]
FIG. 19 shows a neighborhood region of the non-deterministic pixel P1. Since the character image is a two-dimensional image, there are density cross sections in four directions. Binarization processing is performed using information on the character boundary region generated by the four-direction scanner input. First, the first-order differential calculation process 601 by the directional differential operator of FIG. 6 is performed on the multi-value character image I1. The image is subjected to raster scanning in the horizontal direction, the target non-deterministic pixel P is searched, and a primary differential value (gradient of the density cross section) is obtained by applying a filter weighted in each direction around the pixel P. calculate. This is calculated as a feature value for examining in which direction the binarization result shown later is more effective. The greater the gradient of the density cross section, the clearer where the non-deterministic pixels P are distributed in the character boundary region, which is considered effective for binarization determination.
[0063]
After the first-order differential calculation process 601, a binarization result acquisition process 602 in a region near the non-deterministic pixel P1 in the + 45 ° direction is performed. A non-deterministic pixel P1 existing in the direction from the target pixel P to the definite pixel P2 is set as a neighboring area, and a local threshold T2 is set for each neighboring area. FIG. 20 shows the density distribution of the non-deterministic pixel P1 that is determined as a neighboring region distributed in the + 45 ° direction of FIG. A threshold value T2 of the neighborhood region is set for each direction and binarization processing is performed.
[0064]
Thereafter, similar to the binarization result acquisition process 602 in the vicinity region of the non-deterministic pixel P1 in the + 45 ° direction, the same process is performed by changing the direction. Binarization result acquisition process 603 in the vicinity area of the non-deterministic pixel P1 in the horizontal direction, binarization result acquisition process 604 in the vicinity area of the non-deterministic pixel P1 in the -45 ° direction, non-deterministic pixel in the vertical direction The binarization result acquisition process 605 in the vicinity of P1 is sequentially performed to obtain binarization results for each of the four directions. The final binarization result is obtained by determining the gradient of the concentration cross section, which is the above-mentioned first-order differential value, by the voting process 606 in the binarization result for each direction. The above process is performed on all the non-deterministic pixels P1 and selected pixels. This voting process is, for example, that the horizontal binarization result is text and the primary differential value is 50, the vertical binarization result is the background, the primary differential value is 100, and the binarization result in the + 45 ° direction is the background. Suppose that the primary differential value is 10, the binarization result in the −45 ° direction is text, and the primary differential value is 35. In this case, the primary differential value in the horizontal direction and the −45 ° direction is voted to 85 for the character, and the primary differential value in the vertical direction and the + 45 ° direction is voted to 110 for the background. The result of pixel binarization is the background.
[0065]
All pixels determined to be non-deterministic pixels P1 are searched by performing horizontal raster scanning and the above processing is performed to obtain a binarized image I4 having only non-deterministic pixel regions.
[0066]
[Fifth embodiment]
FIG. 7 shows the flow of binarization processing for a fixed pixel in the image processing method according to the fifth embodiment of the present invention, compared to the binarization processing 103 other than the character boundary region described in the first embodiment. Show. In the binarization processing of the definite pixel, binarization processing is performed on the deterministic pixel P2 distributed in the area inside the character and the background area using information of the binarized non-deterministic pixel P1 in the vicinity.
[0067]
As the input image, the binarized image I4 of only the region of the non-deterministic pixel P1 in which the character boundary region generated by the scanner input in the non-deterministic pixel binarization processing 202 of the first embodiment is binarized is used.
[0068]
A region used for binarization of the definite pixel P2 is determined by the determination process 701 for the vicinity region of the definite pixel P2. FIG. 21 shows a vicinity region of the definite pixel P2. In this way, eight neighboring pixels of the definite pixel P2 of interest are searched for and determined as neighboring regions.
[0069]
After determining the neighborhood area, a voting process 702 is performed to binarize the definite pixel P2. The number of binarized non-deterministic pixels P1 distributed in locations determined as neighboring areas is examined and voted, and binaries are integrated by being voted more frequently. If the number is the same, the enlargement process 703 of the neighborhood area of the definite pixel P2 is further performed, and the neighborhood area is enlarged as shown in the right figure of FIG. 702 is performed and binarization is performed. Raster scanning is performed in the horizontal direction from the upper left of the target image to search for the definite pixel P2, and the above processing is performed. When only the definite pixel P2 exists in the vicinity region of the definite pixel P2 of interest, the binarization result of the definite pixel P2 that has already been processed is also reflected in the vote and binarized.
[0070]
FIG. 22 shows the result of the above processing for an image in which the character boundary region is binarized and the definite pixel P2 is not binarized. As shown in FIG. 22, the definite pixel P2 surrounded by the background is binarized to the background pixel, and the definite pixel P2 surrounded by the character pixel is binarized to the character pixel.
[0071]
The above-described definite pixel binarization process is performed on all the pixels determined to be definite pixels P1 to obtain a binary character image (before character boundary region correction) I5.
[0072]
[Sixth embodiment]
FIG. 8 shows the flow of character boundary region correction processing in the image processing method according to the sixth embodiment of the present invention, compared to the character boundary region correction processing 104 described in the first embodiment. In this character boundary region correction process, the unevenness generated in the character boundary region is removed by performing a process of filling it by smoothing.
[0073]
As the input image, a binary character image (before correction of the character boundary region) I5 in which all pixels after the binarization processing 203 of the definite pixel P2 are binarized is used.
[0074]
As shown in FIG. 13, the character image is raster-scanned twice in the x (horizontal) direction, and background pixel filling processing 801 distributed between successive regions of the character image in the horizontal direction is performed. The character pixel filling processing 802 distributed between the continuous regions of the background pixels is performed. In the first x-direction raster scan, a character pixel is searched, and after a continuous area of character pixels is extracted, it is searched whether a background pixel continues in that direction. If it continues, the continuous area of the next character pixel is searched to the last pixel in the x direction. If not, since the character image is distributed again after the background pixel, the background pixel is filled in the character pixel and the continuous region of the character pixel is connected. This process is performed for all x directions. Next, paying attention to the background pixels, the same processing as described above is performed on the continuous region of the background pixels, and the processing of connecting the continuous regions of the background pixel is performed in all the x directions.
[0075]
Similarly, in the y (vertical) direction, raster scanning is performed twice, and the filling process 803 between continuous regions of vertical character pixels and the filling of character pixels distributed between continuous regions of background pixels in the vertical direction are performed. Processing 804 is performed.
[0076]
FIG. 23 shows an example of the smoothing process of the character boundary region. 23 (a) and 23 (c), the unevenness of the boundary area that appears to be noise existing in the boundary area of the character is removed by smoothing, and characters such as (b) and (d) in FIG. 23 are removed. A binary character image I3 in which the boundary region is corrected is output. Focusing on the binarized character image boundary, the unevenness of the boundary region generated in the binarization process due to the influence of low resolution is extracted, smoothed and removed. If a continuous area of character pixels in the vertical direction is detected and the interval between the continuous areas is equal to or smaller than a threshold value, the unevenness of the character boundary area is filled by filling the pixels with the character pixels. Similar processing is performed on a continuous region of character pixels in the horizontal direction. When the interval between the continuous regions is equal to or smaller than a threshold value, the unevenness of the character boundary region is filled by filling the pixel with the character pixel. In addition, the same process is performed on the vertical and horizontal continuous areas of the background pixels to fill the irregularities in the boundary area of the character image. FIG. 24 shows an execution example of the smoothing process.
[0077]
[Seventh embodiment]
FIG. 9 shows processing for switching binarization processing for each region in the image processing method according to the seventh embodiment of the present invention. In the separation process of the non-deterministic pixel P1, an area is assumed from the density cross section of the character image, and the non-deterministic pixel P1 distributed in the character boundary area generated by the scanner input is extracted. First, in select processing 901, regions are separated using feature quantities such as primary differentiation and density values. Then, different binarization processes 902 and 903 are performed for each region.
[0078]
FIG. 12 shows a density distribution diagram of an image subjected to the non-deterministic pixel separation process 201 of the second embodiment. The non-deterministic pixel P1 that is the character boundary region is extracted from the first derivative and the density information in this way, and the local threshold value T2 in the method of the seventh embodiment is applied to the pixel that is greatly affected by the scanner input. Set and binarize.
[0079]
FIG. 25 shows a distribution diagram of the non-deterministic pixel P1 and the definite pixel P2 in the density cross section. From this figure, it can be seen that the non-deterministic pixel P1 is distributed in the character boundary region corresponding to the character boundary region, and the definite pixel P2 is distributed in the region where the density change is flat. By separating the pixels distributed in the character boundary region generated by the scanner input from the pixels distributed in the region other than the character boundary region, each density cross section for each non-deterministic pixel P1 distributed in the character boundary region generated by the scanner input A threshold value T2 is set to the binarization process.
[0080]
After binarization processing by each binarization technique, the result integration processing 904 integrates the results separated and binarized in the selection processing 901, and the result is the final binary character image. Output as I3.
[0081]
[Eighth embodiment]
FIG. 10 shows the flow of binarization processing for non-deterministic pixels in the image processing method according to the eighth embodiment of the present invention compared to the binarization processing 102 for the character boundary region described in the first embodiment. Show. In this binarization processing of non-deterministic pixels, local binarization processing is performed on the pixels in the character boundary region based on the density cross-section information. For the input of this processing, a multi-value character image I1 obtained by scanner input and a separated image I2 of the character boundary region are used. The fourth embodiment is different from the binarization result acquisition process 1002 based on the new threshold value in the vicinity region of the non-deterministic pixel P1 in the + 45 ° direction to the second threshold value based on the new threshold value in the vicinity region of the non-deterministic pixel P1 in the vertical direction. The threshold setting method in the process up to the valueization result acquisition process 1005 is different.
[0082]
In the binarization result acquisition process 1002 based on a new threshold value in the vicinity region of the non-deterministic pixel in the + 45 ° direction, the non-deterministic pixel P1 distributed in that direction existing in the vicinity region to the position of the definite pixel P2 As an area, a local threshold is set for each of the neighboring areas. FIG. 26 shows a threshold shift for preventing the character from curling or collapsing. In the fourth embodiment, the density value of the non-deterministic pixel P1 determined as the neighborhood area is set as the area, and the average value of the maximum and minimum densities is set as the threshold value T2. On the other hand, in the eighth embodiment, the threshold value T2 is shifted using the shape of the density distribution of the non-deterministic pixel P1 in which the target pixel is distributed so that small peaks and valleys of the density cross section can be reproduced. , Prevent curling and crushing of letters. As shown in (a) of FIG. 26, when the shape of the density cross section is distributed so as to draw a mountain, the threshold value is lowered in order to prevent the character from being blurred. As in (b), when the shape of the density cross section is distributed so as to draw a valley, the threshold value is raised to prevent the characters from being crushed. In the case of other shapes, monotonous increase or decrease, the threshold value is not shifted.
[0083]
In the voting process 1006, a process similar to that of the fourth embodiment is performed to obtain a final binarization result. The non-deterministic pixel P1 is searched in the horizontal direction of the image, and the above processing is performed on all non-deterministic pixels P1 to obtain a binarized image I4 having only the non-deterministic pixel P1 region.
[0084]
The embodiment of the present invention has been described above. The above-described embodiment shows an example of a preferred embodiment of the present invention, and the present invention is not limited thereto, and various modifications can be made without departing from the scope of the invention. .
[0085]
【The invention's effect】
As is apparent from the above description, according to the invention described in claim 1, the conventional binarization method extracts a character boundary region generated by a scanner input having a high probability of erroneous binarization determination, and performs binarization. Since the binarization is performed by judging from the density cross-section information where the pixels are distributed in the character image, the image quality of the character image after output can be improved. In addition, the OCR recognition rate can be improved.
[0086]
According to the second aspect of the present invention, the character boundary region generated by the scanner input is separated from the character image. In the conventional binarization method, it is possible to change the binarization process for a character boundary region generated by a scanner input having a high probability of erroneous binarization determination.
[0087]
According to the third aspect of the present invention, the character boundary region generated by the scanner input from the character image is efficiently separated by using the LBG algorithm. The LBG algorithm can divide the region by the number of initial points set first. In the conventional binarization method, it is possible to change the binarization process for a character boundary region generated by a scanner input having a high probability of erroneous binarization determination.
[0088]
According to the fourth aspect of the present invention, it is possible to set a local threshold value using the information of one density section for the character boundary region generated by the scanner input. Since the threshold value is set from one density section, it is possible to extract a background region surrounded by a thick character line or to faithfully reproduce a thin character line. In addition, the OCR recognition rate is improved.
[0089]
According to the fifth aspect of the present invention, in the area other than the character boundary area caused by the scanner input, the background is erroneously determined inside the character, or the character area does not easily exist in the background area. The image quality of the character image is improved, and the recognition rate is improved in the OCR that is sensitive to the noise of the granular character image in the character region.
[0090]
According to the sixth aspect of the present invention, the unevenness of the character boundary area generated after the binarization process is removed by smoothing by the hole filling process, so that the image quality of the character image is improved and the character boundary area The recognition rate is improved in OCR sensitive to unevenness.
[0091]
According to the seventh aspect of the invention, it is possible to perform a local binarization process in which a local threshold value is changed using information of one density section for each character boundary region generated by scanner input. Since a threshold value is set from one density section, it is possible to extract a background region surrounded by a thick character line, or to faithfully reproduce a thin character line. In addition, the OCR recognition rate is improved.
[0092]
According to the eighth aspect of the present invention, it is possible to set a local threshold value using character density cross section and information on the shape of a character boundary region generated by scanner input. Since the threshold value is set by changing the character density section and the shape information, it is possible to extract a background region surrounded by a thick character line and faithfully reproduce a thin character line. In addition, the OCR recognition rate is improved.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating an image processing method according to a first embodiment of the present invention.
FIG. 2 is a diagram showing a flow of binarization processing in the image processing method according to the first embodiment of the present invention.
FIG. 3 is a diagram illustrating another method of binarization processing of the image processing method according to the first embodiment of the present invention.
FIG. 4 is a diagram illustrating non-deterministic pixel separation processing of the image processing method according to the second embodiment of the present invention.
FIG. 5 is a diagram illustrating non-deterministic pixel separation processing of an image processing method according to a third embodiment of the present invention.
FIG. 6 is a diagram illustrating a binarization process of an image processing method according to a fourth embodiment of the present invention.
FIG. 7 is a diagram illustrating binarization processing of a definite pixel in an image processing method according to a fifth embodiment of the present invention.
FIG. 8 is a diagram illustrating a character boundary region correction process in an image processing method according to a sixth embodiment of the present invention.
FIG. 9 is a diagram illustrating binarization processing switching processing of an image processing method according to a seventh embodiment of the present invention.
FIG. 10 is a diagram illustrating non-deterministic pixel binarization processing of an image processing method according to an eighth embodiment of the present invention.
FIG. 11 is a diagram illustrating an example of binarization processing using a simple threshold A and a local threshold B;
FIG. 12 is a diagram illustrating an image that has been subjected to non-deterministic pixel separation processing;
FIG. 13 is a diagram showing removal of irregularities in a character boundary region.
FIG. 14 is a diagram showing a density cross section of a character image.
FIG. 15 is a diagram showing a distribution of primary differentiation and density value.
FIG. 16 is a diagram illustrating a separation process of a non-deterministic pixel P1 by an LBG algorithm.
FIG. 17 is a diagram illustrating a distance calculation formula between a center point and a distribution point.
FIG. 18 is a diagram illustrating an example of a density cross section of a character image.
FIG. 19 is a diagram illustrating a vicinity region of a non-deterministic pixel P1.
FIG. 20 is a diagram illustrating a density distribution of a non-deterministic pixel P1 that is determined to be a neighboring region distributed in a + 45 ° direction.
FIG. 21 is a diagram illustrating a vicinity region of a definite pixel P2.
FIG. 22 is a diagram illustrating an execution example of binarization processing of a definite pixel P2.
FIG. 23 is a diagram illustrating a smoothing process for a character boundary region.
FIG. 24 is a diagram illustrating an execution example of a smoothing process for a character boundary region;
FIG. 25 is a diagram showing a distribution of non-deterministic pixels P1 and definite pixels P2 in a density section.
FIG. 26 is a diagram showing a shift of a threshold value for preventing character curling and crushing.
FIG. 27 is a diagram illustrating a configuration of a general character reading device.
FIG. 28 is a diagram illustrating unevenness of a character boundary region after binarization processing due to a difference in resolution.
[Explanation of symbols]
I1 Scanner input multi-valued character image
I2 Character border region separation image
I3 Binary character image
I4 Binary image of non-deterministic pixel area only
I5 Binary character image (before correction of character boundary area)
P1 non-deterministic pixel
P2 definite pixel
T1 threshold used for non-deterministic pixel separation processing
T2 Threshold value in the vicinity of non-deterministic pixels
1 Scanner section
2 Image storage
3 CPU (image processing unit)
4 OCR (character recognition processing unit)

Claims

Scanner means for digitizing character information formed on a paper surface and capturing it as multi-value image information, arithmetic processing means for performing image processing arithmetic, and an image for storing the image information and a result of image processing by the arithmetic processing means An image processing apparatus comprising: a storage unit; a separation unit that separates a character boundary region generated by input from the scanner unit; a first binarization unit that binarizes the character boundary region; and the character boundary. An image processing apparatus comprising: a second binarization unit that binarizes a region other than the region; and a correction unit that smoothes and corrects the unevenness of the character boundary region after binarization processing,
The first binarization means sets non-deterministic pixels distributed from the target pixel to the deterministic pixel in the horizontal, vertical, and ± 45 degree directions with respect to the character boundary area as a neighboring area, and sets the average density value of the neighboring area. Based on the threshold value set according to the binarization processing to the character and the background,
Horizontal, vertical, relative to the binarization result of each ± 45 ° direction, by weighting the first-order derivative value a is extracted result based on the concentration gradient of the density value and the character image of the neighboring region, the binarization result is character Compare the sum of the first derivative values in all directions with the sum of the first derivative values in all directions where the binarization result is the background, and the larger one is used as the character or background as the final binarization result. An image processing apparatus that performs a voting process to decide.

The image processing apparatus according to claim 1, wherein the separation unit separates the character boundary region using information of a primary differential value and a density value as a feature amount.

The image processing apparatus according to claim 1, wherein the separation unit separates the character boundary region by an LBG algorithm using a first-order differential-density plane of an input image.

The image processing apparatus according to claim 1, wherein the first binarization unit binarizes the character boundary region using character density section information generated by input by the scanner unit.

The image processing apparatus according to claim 1, wherein the second binarization unit performs binarization processing using a binarization processing result of the adjacent character boundary region.

The image processing apparatus according to claim 1, wherein the correction unit performs smoothing by detecting and filling in irregularities generated in the character boundary region of the image after the binarization processing.

The image processing apparatus according to claim 1, wherein the character boundary area is separated and binarization processing is switched for each separated area.

2. The first binarization unit binarizes the character boundary region using a character density section generated by input by the scanner unit and information on the shape thereof. Image processing device.

A separation step of separating a character boundary region generated by the scanner input, a first binarization step of binarizing the separated character boundary region, and the character boundary region; An image processing method comprising: a second binarization step for binarizing a region other than the correction step; and a correction step for smoothing and correcting irregularities in the character boundary region after the binarization processing ,
In the first binarization step, horizontal with respect to the character boundary region, vertical, non-deterministic pixels distributed from the pixel of interest for each ± 45 ° direction to determine the pixel as the neighboring region, the average density value of the neighboring region Based on the threshold value set according to the binarization processing to the character and the background,
Horizontal, vertical, relative to the binarization result of each ± 45 ° direction, by weighting the first-order derivative value a is extracted result based on the concentration gradient of the density value and the character image of the neighboring region, the binarization result is character Compare the sum of the first derivative values in all directions with the sum of the first derivative values in all directions where the binarization result is the background, and the larger one is used as the character or background as the final binarization result. An image processing method characterized by performing a voting process to decide.

The image processing method according to claim 9, wherein the separation step separates the character boundary region using information of a primary differential value and a density value as feature amounts.

The image processing method according to claim 9, wherein the separating step separates the character boundary region by an LBG algorithm using a first-order differential-density plane of the input image.

The image processing method according to claim 9, wherein the first binarization step binarizes the character boundary region using character density section information generated by the scanner input.

The image processing method according to claim 9, wherein the second binarization step performs binarization processing using a binarization processing result of the adjacent character boundary region.

The image processing method according to claim 9, wherein the correcting step performs smoothing by detecting and filling unevenness generated in the character boundary region of the image after the binarization processing.

The image processing method according to claim 9, wherein the character boundary region is separated, and binarization processing is switched for each separated region.

The image processing according to claim 9, wherein the first binarization step binarizes the character boundary region by using character density cross sections generated by the scanner input and information on the shape thereof. Method.