CN102968799B - Integral image-based quick ACCA-CFAR SAR (Automatic Censored Cell Averaging-Constant False Alarm Rate Synthetic Aperture Radar) image target detection method - Google Patents

Abstract

The invention provides an integral image-based quick ACCA-CFAR SAR (Automatic Censored Cell Averaging-Constant False Alarm Rate Synthetic Aperture Radar) image target detection method, comprising the following steps of: (1) providing a G0 distribution-based self-adaptive global threshold CFAR pre-segmentation algorithm used for generating a target index matrix by combining the statistical property of data; (2) providing an integral image-based G0 distribution statistical parameter quick estimation method, wherein the statistical parameter can be calculated through simple operations such as addition and subtraction once 2-order and 4-prder integral images of an original image are obtained during the implementation of the method; and (3) giving out a basic implementation process of the ACCA-CFAR SAR image target detection method. Through the integral image-based G0 distribution statistical parameter quick estimation strategy provided by the invention, the time efficiency of the method can be greatly improved and the time complexity of the method is irrelevant to the size of a sliding window; and the requirement of the existing automatic target recognition (ATR) system on the treatment of large-scene data can be met to a great extent.

Description

A kind of quick A CCA-CFAR SAR image object detection method based on integral image

Technical field

The present invention relates to SAR image interpretation technical field, be specifically related to a kind of quick A CCA-CFARSAR image object detection method based on integral image.

Background technology

High resolving power, large scene synthetic-aperture radar (Synthetic Aperture Radar, SAR) emerging in large numbers of image, provide possibility for being applied even more extensively SAR image, brought new challenge to SAR image interpretation technology simultaneously, in being adapted to before some, low resolution, the treatment technology of little scene SAR image is no longer applicable.As one of crucial SAR image interpretation technology, target detection has a significant impact performance and the efficiency tool of the subsequent treatment such as feature extraction, target identification and classification.At present in existing certain development aspect this, CFAR (Constant False Alarm Rate, CFAR) detecting is the object detection method being wherein most widely used, its ultimate principle is: also adjust successively thresholding according near the energy of reference unit estimated background clutter detecting unit, with make false-alarm probability constant (referring to document [1] Xu little Jian, Huang Peikang, " radar system and information processing thereof; " Electronic Industry Press, 2010.).

But, traditional CFAR operator, be cell-average (Cell Averaging, CA)-CFAR(is referring to document [2] L.M.Novak, G.J.Owirka, W.S.Brower, and A.L.Weaver, " The automatic target-recognition system inSAIP " Linc.Lab.J., Vol.10, No.2,1997, pp.187-202.), suppose background clutter Gaussian distributed, the single goal being only applicable in local uniform background clutter detects, and in the time that background clutter is non-homogeneous or comprise multiple goal, it detects performance and sharply declines.In order to meet the demand of practical engineering application, in the urgent need to obtain high resolving power, data content complexity and the huge feature of capacity of data for novel sensor, the SAR imaging characteristic difference of combining target and background clutter, research is applicable to the quick CFAR object detection method of various complex scenes.

The technical scheme of prior art one:

In order to meet the application demand for the treatment of S AR image complex scene, there is researcher to consider the GO of each tool relative merits (Greatest Of)-CFAR(referring to document [3] V.G., Hansen, " Constant false alarm rate processing insearch radars, " In Proceedings of the IEEE1973International Radar Conference, London, 1973, pp.325-332.), SO (Smallest Of)-CFAR(is referring to document [4] G.V.Trunk, " Range resolution of targets usingautomatic detectors, " IEEE Transactions on Aerospace and Electronic Systems, AES-14, Sept.1978, etc. pp.750-755.) basic operator combines according to certain criterion.Wherein, the most representative VI (VariabilityIndex)-CFAR(is referring to document [5] M.E.Smith and P.K.Varshney, " Intelligent CFAR processor based on data variability, " IEEE Trans.Aerosp.Electron.Syst., Vol.36, No.3, Jul.2000, pp.837 – 847.) operator, according to index value and test of hypothesis average ratio based on second-order statistics feature, judge and select CA-CFAR, one of GO-CFAR and SO-CFAR operator carry out target detection, therefore have three's advantage concurrently, be applicable to process even scene, and comprise multiple goal, the complex scene data of clutter edge etc.

Adopt similar thinking, document [6] (G.Gao, L.Liu, L.J.Zhao, G.T.Shi, et al., " An adaptive and fastCFAR algorithm based on automatic censoring for target detection in high-resolution SARimages; " IEEE Trans.Geosci.Remote Sens., Vol.47, No.6,2009, pp.1685-1697.) a kind of quick CFAR object detection method based on automatic retrieval (Automatic Censoring, AC) is proposed.As Fig. 1 center 1. as shown in, the method generates target index matrix by a kind of pre-segmentation algorithm of based target degree of confidence and carrys out the candidate target pixel in mark backdrop window, so that they are rejected from reference unit, improve the applicability of operator for multiple goal scene, the light grey square frame in figure represents target index pixel.How to calculate quickly and efficiently the key that object pixel index value is these class methods.

The shortcoming of prior art one:

Although by multiple basic operator dominance complementations can be improved to the applicability of CFAR algorithm for complex scene, still there is following shortcoming in these class methods:

(1) the VI-CFAR operator in document [5] is not deeply considered the statistical property of time performance and the background clutter of algorithm.This operator adopts the index value computing method based on local window processing, and its hypothesis background clutter Gaussian distributed, determines index value by average and the variance of estimating reference unit.Calculation of complex, time efficiency are lower, and selected empirical model is not suitable for non-homogeneous data.

(2) the global threshold pre-segmentation algorithm of the based target degree of confidence that document [6] adopts is not considered the statistical property of data, its performance only depends on objective degrees of confidence, be object pixel number and the ratio of the total number of pixels of image, thereby the performance of algorithm is more responsive to the value of this parameter.

The technical scheme of prior art two:

The statistical model that more can accurately reflect background clutter statistical property is introduced CFAR operator by part researcher, to improve the target detection ability of algorithm under complex scene.Wherein, it is that a kind of compound Gauss model of the SAR of meeting scattering mechanism is (referring to document [7] E.Jakeman and P.N.Pusey that K-distributes, " A model for non-Rayleigh sea echo, " IEEE Trans.Antennas Propagat., Vol.AP-24, 1976, and document [8] Hao Chengpeng pp.806-814., Hou Chaohuan, DP-CFAR detecting device [J] under a kind of K-Distribution Clutter background, electronics and information journal, Vol.29, No.3, 2007, pp.756-759.), it is because having extra large clutter, forest land, the ability of the non-homogeneous data modeling such as farmland and more concerned, on the other hand, G ⁰distribute (referring to document [9] A.C.Frery, H.J.Muller, C.C.F.Yanasse, and S.J.S.Sant ' Anna, " A model forextremely heterogeneous clutter, " IEEE Trans.Geosci.Remote Sens., Vol.35, No.3, May1997, pp.648-659.) have the ability to the clutter region modeling under extensive uniformity coefficient variation, stronger model compatibility and calculation of parameter are simple.As Fig. 1 center 2. as shown in, this statistical model is applied to CFAR operator by document [6], fully utilizes this model and be applicable to, to equal, the non-homogeneous and characteristic of non-homogeneous scene data modeling extremely, ensure the ability of algorithm process complex scene data.

The shortcoming of prior art two:

Although it is widely used statistical model aspect the research of CFAR operator that K-distributes, along with the raising of SAR image resolution ratio and the increase of content complexity, still there is following shortcoming in this class algorithm:

(1) the K-extreme Nonuniform Domain Simulation of Reservoir modelings such as not being suitable for open marine site to the city in High Resolution SAR Images, inshore marine site or high sea condition that distributes.Fig. 2 and Fig. 3 taking the interested marine site that intercepts from 3m-resolution, large scene SAR image and inshore city (ROI) as test data, have contrasted K-and have distributed and G respectively ⁰distribution range model ( distribute) modeling result.Can find out, distribute and be all better than to a certain extent K-distribution for the modeling result of these two groups of data, especially for non-homogeneous data of extreme such as inshore cities, the performance that K-distributes sharply declines, and distribute and still can obtain good fitting result.

(2) compare and G ⁰distribute, the statistical parameter that K-distributes and the equal more complicated of calculating of local CFAR threshold value.

The technical scheme of prior art three:

In order to meet real-time automatic target identification (Automatic Target Recognition, ATR) system and to process the demand of large scene data, the time efficiency that improves CFAR operator is another problem demanding prompt solution.But, still be in the junior stage in the research aspect this at present, the simplest accelerated method is larger moving window scanning step to be set (referring to document [10] C.H.Jung, W.Y.Song, S.H.Rho, et al., " Double-step fast CFAR scheme for multiple target detectionin high resolution SAR images; " IEEE, 2010, pp.1172-1175.); Document [6] proposes a kind of G ⁰distribution statistics parameter is estimated strategy fast.Consider two-parameter moving window as shown in Figure 4, the size that wherein h is moving window, r represents the width of backdrop window, protecting window is of a size of h-2r.There are most of reference units in conjunction with two of left and right (or upper and lower) adjacent window apertures of CFAR operator identical; only backdrop window and the different feature of the individual reference unit of protecting window boundary h+ (h-2r); taking scan method from left to right as example; can be according to strategy as shown in Figure 5; avoid some repetitive operations, improve the time performance of algorithm.

${{\mathrm{\μ}}^2}_{\mathrm{right}}=\frac{N_C\·{{\mathrm{\μ}}^2}_{\mathrm{left}}-(\underset{i=1}{\overset{h}{\mathrm{\Σ}}}{x}_{l,i}^2+\underset{i=1}{\overset{h-2r}{\mathrm{\Σ}}}{x}_{r,i}^2)+\left(\underset{i=1}{\overset{h}{\mathrm{\Σ}}}{x}_{r,i}^2+\underset{i=1}{\overset{h-2r}{\mathrm{\Σ}}}{x}_{l,i}^2\right)}{N_C}---\left(A\right)$

${\mathrm{\μ}}_{\mathrm{right}}^4=\frac{N_C\·{\mathrm{\μ}}_{\mathrm{left}}^2-(\underset{i=1}{\overset{h}{\mathrm{\Σ}}}{x}_{l,i}^4+\underset{i=1}{\overset{h-2r}{\mathrm{\Σ}}}{x}_{r,i}^4)+(\underset{i=1}{\overset{h}{\mathrm{\Σ}}}{x}_{r,i}^4+\underset{i=1}{\overset{h-2r}{\mathrm{\Σ}}}{x}_{l,i}^4)}{N_C}---\left(B\right)$

In formula, with with represent respectively the 2-rank of effective reference unit in left and right backdrop window, 4-rank sample moment; x _l,iand x _r,icorrespond respectively to the reference unit at left and right window edge place; " subtract " item corresponding to the Dark grey dash area in Fig. 5, " adding ", Xiang Ze was corresponding to light grey dash area.

Suppose all N in whole backdrop window _cindividual pixel is background pixel, and in document [6], complexity computing time of fast method is (N ²-1) (3h+12)+3N _c+ 6, the time complexity of traditional C FAR operator is N ²(3N _c+ 2).Therefore the fast method in document [6] can be down to time complexity the 1/4r of traditional C FAR operator.

The shortcoming of prior art three:

Although the express statistic parameter estimation strategy in document [6] is by avoiding some repetitive operations, improve the time efficiency of algorithm on largely.But described in document [6], when moving window parameter is set to h=71, r=20, picture size is 1375 × 1880 o'clock, be still 40.0471s the working time of this algorithm, and known this algorithm still can not meet the application demand of large scene data processing and real-time ATR system far away.

Summary of the invention

Technical matters to be solved by this invention is: for the deficiency of existing CFAR object detection method, propose a kind of new quick A CCA-CFAR SAR image object detection method, comprising: (1), in conjunction with the statistical property of data, proposes a kind of based on G ⁰the self-adaptation global threshold CFAR pre-segmentation algorithm distributing, for generating target index matrix.(2) a kind of G based on integral image is proposed ⁰distribution statistics parameter method for quick estimating, in the method implementation procedure, once try to achieve 2-rank, the 4-rank integral image of original image, can try to achieve statistical parameter by the computing such as simply adding, subtract, thereby greatly improve the time efficiency of algorithm; (3) provided the basic realization flow of this ACCA-CFAR algorithm of target detection.In addition, because the algorithm of pre-segmentation part and succeeding target test section is based on G ⁰the CFAR operator distributing, just the former is based on global threshold, the latter is local threshold, thereby both implementation methods are similar, thus the entirety that has reduced algorithm realizes difficulty.

The technical solution adopted in the present invention is as follows: a kind of quick A CCA-CFAR SAR image object detection method based on integral image, and the method is specific as follows:

(1) in conjunction with the statistical property of data, utilize based on G ⁰the self-adaptation global threshold CFAR pre-segmentation algorithm distributing generates target index matrix;

(2) adopt the G based on integral image ⁰distribution statistics parameter method for quick estimating, wherein, after trying to achieve the 2-rank, 4-rank integral image of original image, can try to achieve statistical parameter by the computing such as simply adding, subtract;

(3), by the quick A CCA-CFAR algorithm based on integral image, realize SAR image object and detect.

Wherein, described utilization is based on G ⁰the self-adaptation global threshold CFAR pre-segmentation algorithm distributing generates target index matrix, and its concrete steps are as follows:

Step 11: invariable false alerting p is set _fa, two-parameter moving window is extended to entire image, and protecting window is set is of a size of 0;

Step 12: the reference unit in backdrop window is used for to G ⁰distribution range modeling statistics parameter estimation,

In formula, n presentation video look number; μ ², μ ⁴the 2-rank, the 4-rank sample moment that represent respectively reference unit, their account form is:

${\mathrm{\μ}}^m=\frac{1}{N_s}\underset{t=1}{\overset{N_s}{\mathrm{\Σ}}}{\left|z\left(t\right)\right|}^m---\left(3\right)$

In formula, m represents exponent number, and z (t) is each reference unit, N _srepresent reference unit number;

Step 13: according to CFAR operator ultimate principle formula, calculate global threshold T _g,

$1-p_{\mathrm{fa}}={\∫}_0^{T_g}{f}_{G^0}\left(z\right)\mathrm{dz}---\left(4\right)$

In formula, represent G ⁰the probability density function of distribution range model;

$f_{G^0}\left(z\right)=\frac{2n^n\mathrm{\Γ}(n-\mathrm{\α}){\mathrm{\γ}}^{-\mathrm{\α}}{z}^{2n-1}}{\mathrm{\Γ}\left(n\right)\mathrm{\Γ}(-\mathrm{\α}){(\mathrm{\γ}+n{z}^2)}^{n-\mathrm{\α}}}---\left(5\right)$

In formula, alpha, gamma is respectively shape and scale parameter ,-α > 0, γ > 0; Work as n=1, when haplopia data, the analytic solution of formula (4) are:

For looking data, because formula (4) is without analytic solution, adopt dichotomy to determine more

Step 14: according to the criterion shown in formula (7), in original image to each test pixel I _tadjudicate,

Wherein H ₁and H ₀represent respectively object pixel hypothesis and background pixel hypothesis, thus the target index matrix M shown in production (8):

$m(x,y)=\left\{\begin{array}{c}1,H_1\\ 0,H_0\end{array}\right.---\left(8\right)$

In formula, (x, y) represents the coordinate position of pixel.

Wherein, the G of described employing based on integral image ⁰distribution statistics parameter method for quick estimating, its concrete steps are as follows:

Step 21: by repeatedly carrying out following two groups of computings, the m-rank image i of an original image i of scanning (x, y) ^m(x, y), generates its m-rank integral image ii ^m(x, y),

s(x,y)=s(x,y-1)+i ^m(x,y) (9)

ii ^m(x,y)=ii ^m(x-1,y)+s(x,y) (10)

In formula, (x, y) represents the coordinate position of current pixel point; wherein i _k()=i () (k=1 ..., m), " dot product " of Π presentation video; S (x, y) represents i ^m(x, y) follows integration, wherein s (x ,-1)=0, ii ^m(1, y)=0; The needs of estimating according to statistical parameter, get m=2,4;

Step 22: the m-rank integral image based on original image, calculate fast the m-rank sample moment of reference unit in any two-parameter moving window,

${\mathrm{\μ}}^m=\frac{1}{N_s}\underset{t=1}{\overset{N_s}{\mathrm{\Σ}}}{\left|z\left(t\right)\right|}^m$ (11)

$=\frac{1}{N_s}\left\{\right[({\mathrm{ii}}_1^m+{\mathrm{ii}}_4^m)-({\mathrm{ii}}_2^m+{\mathrm{ii}}_3^m)]-[({\mathrm{ii}}_{1^{\′}}^m+{\mathrm{ii}}_{4^{\′}}^m)-({\mathrm{ii}}_{2^{\′}}^m+{\mathrm{ii}}_{3^{\′}}^m)\left]\right\}$

In formula, (p=1,2,3,4,1', 2', 3', 4') is the pixel value of p place, summit m-rank integral image; corresponding to pixel value sum in whole moving window, for pixel value sum in protecting window;

Step 23: by 2-rank, 4-rank sample moment substitution formula (1) and (2), can try to achieve G by simple operation ⁰the statistical parameter of model.

Wherein, described pass through the quick A CCA-CFAR algorithm based on integral image, realize SAR image object and detect, its specific implementation step is as follows:

Step 301: adopt based on G ⁰the global threshold CFAR algorithm distributing carries out target pre-segmentation, generates target index matrix; For the ease of follow-up computing, to this index matrix negate, obtain background pixel index matrix, wherein background dot is 1, impact point is 0, for marking all candidate background pixels;

Step 302: by by background pixel index matrix and original image dot product, remove the interference of candidate target pixel, obtain background clutter image, with the reservation background pixel value of original image equidimension, the image of target pixel value zero setting;

Step 303: the m-rank integral image that calculates background clutter image;

Step 304: invariable false alerting and moving window parameter are set, and wherein the size of protecting window should be greater than the size of target to be detected, sliding window size should participate in statistical parameter estimation to ensure abundant background pixel fully greatly;

Step 305: utilize moving window by picture element scan original image, and by m-rank integral image and background pixel index matrix, calculate the m-rank sample moment of background pixel;

Step 306: utilize the m-rank sample moment of background pixel, adopt MoM method to estimate G ⁰the statistical parameter distributing;

Step 307: utilize statistical parameter to calculate CA-CFAR local threshold;

Step 308: by the size of compare test unit and local threshold, judge that whether this unit is candidate target pixel, is to be 1, otherwise is 0;

Step 309: judge whether to continue scanning, if do not scan complete image, skip to step 305, scan next test cell; Otherwise, carry out next step;

Step 310: remove false-alarm by subsequent treatment, merge object pixel region, thereby obtain final target detection result.

Wherein, the subsequent treatment described in step 310 is counting wave filter, morphology processing or target area cluster.

The beneficial effect that technical solution of the present invention is brought is:

Compare and existing CFAR object detection method, the beneficial effect that the quick A CCA-CFAR method based on integral image proposed by the invention is brought embodies in the following areas:

(1) by introducing AC technology and G ⁰distributed model, has the target detection ability in SAR image complex scene;

(2) propose based on G ⁰the self-adaptation global threshold CFAR pre-segmentation algorithm distributing has been considered time efficiency and the data statistics characteristic of algorithm, to the sensitivity of parameter lower and its realize principle and succeeding target test section to realize principle similar, thereby the entirety that has reduced algorithm realizes difficulty;

(3) strategies based on integral image proposing can improve the time efficiency of algorithm greatly, and this fast method makes time complexity and the moving window cache oblivious of algorithm, thereby can be the in the situation that of influence time efficiency not, set according to actual needs the accuracy of enough large sliding window size with the modeling of guarantee clutter statistical characteristics.

Brief description of the drawings

Fig. 1 is the basic ideas of the self-adaptation based on AC in document [6], quick CFAR algorithm of target detection;

Fig. 2 is distribution, K-distribute the modeling result of extra large clutter in High Resolution SAR Images are contrasted; (a) extra large clutter ROI; (b) extra large clutter PDF with distribution, K-distribution PDF contrast;

Fig. 3 is distribution, K-distribute the modeling result of land clutter in High Resolution SAR Images are contrasted; (a) land clutter ROI; (b) land clutter PDF with distribution, K-distribution PDF contrast;

Fig. 4 is two-parameter moving window schematic diagram;

Fig. 5 is G in document [6] ⁰distribution statistics parameter is estimated schematic diagram fast;

Fig. 6 is based on G ⁰the process flow diagram of the self-adaptation global threshold CFAR pre-segmentation algorithm distributing;

Fig. 7 is the G based on integral image ⁰the process flow diagram of distribution statistics parameter method for quick estimating;

Fig. 8 is m-rank integral image schematic diagram;

Fig. 9 is the quick calculating schematic diagram of reference unit m-rank sample moment in the two-parameter moving window based on integral image;

Figure 10 is the process flow diagram of the quick A CCA-CFAR algorithm of target detection based on integral image;

Figure 11 is original image schematic diagram;

Figure 12 is the target detection result of institute of the present invention extracting method, and with document [6] in contrast (a) the target thumbnail of method; (b) the quick A CCA-CFAR target detection result based on integral image; (c) the final target detection result of the quick A CCA-CFAR method based on integral image; (d) the final target detection result of method in document [6].

Embodiment

Further illustrate the present invention below in conjunction with accompanying drawing and instantiation.

The present invention proposes a kind of quick A CCA-CFAR SAR image object detection method based on integration image, specifically comprises the steps:

(1) based on G ⁰the self-adaptation global threshold CFAR pre-segmentation algorithm distributing

Consider on the one hand G ⁰the above-mentioned advantage distributing; On the other hand because the ratio of object pixel in SAR image is generally very little, if utilize mass data to carry out statistical parameter estimation, object pixel to affect meeting very little, and ACCA-CFAR operator itself is not very high to the accuracy requirement of pre-segmentation algorithm.Therefore, one is proposed based on G ⁰the self-adaptation global threshold CFAR pre-segmentation algorithm distributing, for generating target index matrix.As shown in Figure 6, concrete methods of realizing is as follows for this pre-segmentation algorithm flow chart:

Step 1: as shown in Fig. 6 center 1, invariable false alerting p is set _fa, the two-parameter moving window shown in Fig. 4 is extended to entire image, and protecting window is set is of a size of 0.

Step 2: as shown in Fig. 6 center 2, the reference unit in backdrop window is used for to G ⁰distribution range modeling statistics parameter estimation,

In formula, n presentation video look number; μ ², μ ⁴the 2-rank, the 4-rank sample moment that represent respectively reference unit, their account form is:

${\mathrm{\μ}}^m=\frac{1}{N_s}\underset{t=1}{\overset{N_s}{\mathrm{\Σ}}}{\left|z\left(t\right)\right|}^m---\left(3\right)$

In formula, m represents exponent number, and z (t) is each reference unit, N _srepresent reference unit number.Formula (1) and (2) be adopt method of moment (MoM) (referring to document [11] He Zhiguo, Zhou Xiaoguang, ground force and Kuang Guangyao, " a kind of based on G ⁰the quick CFAR detection method of SAR image distributing, " National University of Defense technology's journal, Vol.31, No.1,2009, pp.47-51.) G that obtains ⁰distribution range modeling statistics parameter calculation formula.

Step 3: as shown in Fig. 6 center 3, according to CFAR operator ultimate principle formula, calculate global threshold T _g,

$1-p_{\mathrm{fa}}={\∫}_0^{T_g}{f}_{G^0}\left(z\right)\mathrm{dz}---\left(4\right)$

In formula, represent G ⁰the probability density function (PDF) (referring to document [9]) of distribution range model.

$f_{G^0}\left(z\right)=\frac{2n^n\mathrm{\Γ}(n-\mathrm{\α}){\mathrm{\γ}}^{-\mathrm{\α}}{z}^{2n-1}}{\mathrm{\Γ}\left(n\right)\mathrm{\Γ}(-\mathrm{\α}){(\mathrm{\γ}+n{z}^2)}^{n-\mathrm{\α}}}---\left(5\right)$

In formula, alpha, gamma is respectively shape and scale parameter ,-α > 0, γ > 0.Work as n=1, when haplopia data, the analytic solution of formula (4) are:

For looking data, because formula (4) is without analytic solution, adopt dichotomy to determine more (referring to document [11]).

Step 4: as shown in Fig. 6 center 4, according to the criterion shown in formula (7), in original image to each test pixel I _tadjudicate,

Wherein H ₁and H ₀represent respectively object pixel hypothesis and background pixel hypothesis, thus the target index matrix M shown in production (8).

$m(x,y)=\left\{\begin{array}{c}1,H_1\\ 0,H_0\end{array}\right.---\left(8\right)$

In formula, (x, y) represents the coordinate position of pixel.

(2) G based on integral image ⁰distribution statistics parameter method for quick estimating

From formula (1)-(3), G ⁰the main computing of distribution statistics parameter estimation is the calculating of reference unit 2-rank, 4-rank sample moment, and the main operand of sample moment is " summation ", if therefore can summation operation part be accelerated, improves surely the integral operation speed of algorithm of target detection.Utilize integral image techniques can calculate fast rectangular characteristic (referring to document [12] P.Viola, M.Jones, " Robust real-time face detection; " International Journal of Computer Vision, Vol.52, No.2,2004, pp.137-154.), can be introduced into the quick calculating of sample moment.Fig. 7 is the G based on integral image ⁰distribution statistics parameter is estimated process flow diagram fast, and its concrete methods of realizing is as follows:

Step 1: as shown in Fig. 7 center 1, by repeatedly carrying out following two groups of computings, the m-rank image i of an original image i of scanning (x, y) ^m(x, y), generates its m-rank integral image ii ^m(x, y), Fig. 8 is the schematic diagram of m-rank integral image.

s(x,y)=s(x,y-1)+i ^m(x,y) (9)

ii ^m(x,y)=ii ^m(x-1,y)+s(x,y) (10)

In formula, (x, y) represents the coordinate position of current pixel point; wherein i _k()=i () (k=1 ..., m), " dot product " of Π presentation video; S (x, y) represents i ^m(x, y) follows integration, wherein s (x ,-1)=0, ii ^m(1, y)=0; The needs of estimating according to statistical parameter, get m=2,4.

Step 2: as shown in Fig. 7 center 2, the m-rank integral image in conjunction with Fig. 9 based on original image, calculates the m-rank sample moment of reference unit in any two-parameter moving window (light grey dash area) fast,

${\mathrm{\μ}}^m=\frac{1}{N_s}\underset{t=1}{\overset{N_s}{\mathrm{\Σ}}}{\left|z\left(t\right)\right|}^m$ (11)

$=\frac{1}{N_s}\left\{\right[({\mathrm{ii}}_1^m+{\mathrm{ii}}_4^m)-({\mathrm{ii}}_2^m+{\mathrm{ii}}_3^m)]-[({\mathrm{ii}}_{1^{\′}}^m+{\mathrm{ii}}_{4^{\′}}^m)-({\mathrm{ii}}_{2^{\′}}^m+{\mathrm{ii}}_{3^{\′}}^m)\left]\right\}$

In formula, (p=1,2,3,4,1', 2', 3', 4') is the pixel value of summit p place m-rank integral image in Fig. 9; corresponding to pixel value sum in whole moving window, for pixel value sum in protecting window.

Step 3: as shown in Fig. 7 center 3, by 2-rank, 4-rank sample moment substitution formula (1) and (2), can try to achieve G by simple operation ⁰the statistical parameter of model.

As shown in table 1, complexity computing time of this statistical parameter method for quick estimating is 2 (N-1) ²+ 24N ².In traditional C FAR operator and document [6], complexity computing time of strategies is respectively with the ratio of this value:

${\mathrm{\υ}}_1>\frac{N^2({3N}_C+2)}{2{(N-1)}^2+{24N}^2}>\frac{N^2({3N}_C+2)}{{26N}^2}=\frac{{3N}_C+2}{26}\≈\frac{r(h-r)}{2}---\left(12\right)$

${\mathrm{\υ}}_2>\frac{(N^2-1)(3h+12)+{3N}_C+6}{2{(N-1)}^2+{24N}^2}>\frac{N^2(3h+12)}{{26N}^2}=\frac{3h+2}{26}\≈\frac{h}{9}---\left(13\right)$

Analysis of complexity computing time of table 1 fast method proposed by the invention

Known, the time efficiency of institute of the present invention extracting method is at least the r (h-r)/2 times of traditional double parameters C FAR algorithm, the h/9 of document [6] algorithm of carrying doubly, and the time complexity of institute of the present invention extracting method and the size of window are irrelevant, and the time complexity of other two kinds of methods is closely related with window parameter.For the stability that ensures that statistical parameter is estimated, moving window and backdrop window size must arrange enough large conventionally.Relatively be set to example with the parameter in document [6] for convenient, i.e. h=71, r=20, the time efficiency that now can calculate institute of the present invention extracting method is at least about 510 times of traditional double parameters C FAR algorithm, 8 times of document [6] algorithm of carrying.

(3) basic procedure of the quick A CCA-CFAR SAR image object detection method based on integral image

Figure 10 is the process flow diagram of quick A CCA-CFAR algorithm proposed by the invention, and its specific implementation step is as follows:

Step 1: as shown in Figure 10 center 1a and 1b, adopt based on G ⁰the global threshold CFAR algorithm distributing carries out target pre-segmentation, generates target index matrix; For the ease of follow-up computing, to this index matrix negate, obtain background pixel index matrix, wherein background dot is 1, impact point is 0, for marking all candidate background pixels.

Step 2: as shown in Figure 10 center 2a and 2b, by by background pixel index matrix and original image " dot product ", remove the interference of candidate target pixel, obtain background clutter image, with the reservation background pixel value of original image equidimension, the image of target pixel value zero setting.

Step 3: as shown in Figure 10 center 3, calculate the m-rank integral image of background clutter image.

Step 4: as shown in Figure 10 center 4, invariable false alerting and moving window parameter are set, wherein the size of protecting window should be greater than the size of target to be detected, sliding window size should participate in statistical parameter estimation to ensure abundant background pixel fully greatly.

Step 5: as shown in Figure 10 center 5a and 5b, utilize moving window by picture element scan original image, and by m-rank integral image and background pixel index matrix, calculate the m-rank sample moment of background pixel.

Step 6: as shown in Figure 10 center 6, utilize the m-rank sample moment of background pixel, adopt MoM method to estimate G ⁰the statistical parameter distributing.

Step 7: as shown in Figure 10 center 7, utilize statistical parameter to calculate CA-CFAR local threshold.

Step 8: as shown in Figure 10 center 8, by the size of compare test unit and local threshold, judge that whether this unit is candidate target pixel, is to be 1, otherwise is 0.

Step 9: as shown in Figure 10 center 9, judge whether to continue scanning, if do not scan complete image, skip to step 5, scan next test cell; Otherwise, carry out next step.

Step 10: as shown in Figure 10 center 10, remove false-alarm by subsequent treatment (as: counting wave filter, morphology processing, target area cluster etc.), merge object pixel region, thereby obtain final target detection result.

By a concrete application process that provides for example the quick A CCA-CFAR object detection method based on integral image that this invention proposes, contrast with the fast method that document [6] proposes below simultaneously.

Suppose to have obtained a width open Sea SAR image as shown in figure 11 by certain sensor, existing need judge fast in this image, whether to comprise Ship Target definite their position.These data are single-view picture, and it is of a size of 961*680, and pixel resolution is 3m.

Suppose according to technical requirement, arranged based on G ⁰the invariable false alerting p of distribution self-adaptation global threshold CFAR pre-segmentation algorithm _fa1=10 ^-2; For the ease of contrast, the parameter of the quick A CCA-CFAR target detection based on integral image adopts and setting identical in document [6], i.e. invariable false alerting p _fa2=10 ^-3, sliding window size h=71, r=20; The subsequent treatment parameter wavenumber filter threshold value of falling into a trap is 5, and morphology processing parameter is 1, and in document [6] as a comparison, method adopts identical therewith setting.

Figure 12 has provided the interim result of institute of the present invention extracting method, and has contrasted the final target detection result of method in itself and document [6].Wherein, Figure 12 (a) is for adopting based on G ⁰the target thumbnail that the self-adaptation global threshold CFAR pre-segmentation algorithm distributing generates, it has detected most of real goal candidate region in the situation that false alarm rate is very little; Figure 12 (b) is the quick A CCA-CFAR target detection result based on integral image, in order to remove discrete false-alarm pixel wherein, has obtained the final target detection result as shown in Figure 12 (c) by counting filtering, morphology processing; Figure 12 (d) is the final target detection result of method in document [6].Contrast Figure 12 (c) and Figure 12 (d), be easily shown in: both testing results are suitable, all effectively detected three naval vessels.

Aspect algorithm time efficiency, for the ease of relatively, in table 2, contrast the working time of two kinds of methods.Yi Zhi: be about 10 times of fast method in document [6] working time of institute of the present invention extracting method, this and aforementioned analysis are basically identical.On the other hand, the working time of two kinds of method pre-segmentation parts is based on quite, but the former is because having considered the statistical property of data, accuracy of detection is higher, to the sensitivity of parameter lower and its realize principle and follow-up CFAR algorithm basically identical, thereby the entirety that has reduced algorithm realizes difficulty.

The time efficiency of method contrast in table 2 this invention institute's extracting method and document [6]

Also can complete goal of the invention by following replacement scheme:

(1) the present invention is applicable to the target detection problems in the multiple SAR image scene such as marine site, land clutter;

(2) the pre-segmentation algorithm of generation target index matrix is replaceable is the global threshold pre-segmentation algorithm based on statistical models such as Gaussian distribution, K-distributions, and other adaptive thresholding algorithm;

(3) strategies based on integral image can be applied to other statistical model in CFAR operator, and the MoM method that for example K-distributes, Pareto distributes and fractional order method of moment (MoFM) statistical parameter are estimated.

Claims (1)

1. the quick A CCA-CFAR SAR image object detection method based on integral image, is characterized in that, the method is specific as follows:

(1) in conjunction with the statistical property of data, utilize based on G ⁰the self-adaptation global threshold CFAR pre-segmentation algorithm distributing generates target index matrix;

Wherein, described utilization is based on G ⁰the self-adaptation global threshold CFAR pre-segmentation algorithm distributing generates target index matrix, and concrete steps are as follows:

Step 11: invariable false alerting p is set _fa, two-parameter moving window is extended to entire image, and protecting window is set is of a size of 0;

Step 12: the reference unit in backdrop window is used for to G ⁰distribution range modeling statistics parameter estimation,

In formula, n presentation video look number; μ ², μ ⁴the 2-rank, the 4-rank sample moment that represent respectively reference unit, their account form is:

${\mathrm{\μ}}^m=\frac{1}{N_s}\underset{t=1}{\overset{N_s}{\mathrm{\Σ}}}{\left|z\left(t\right)\right|}^m---\left(3\right)$

In formula, m represents exponent number, and z (t) is each reference unit, N _srepresent reference unit number;

Step 13: according to CFAR operator ultimate principle formula, calculate global threshold T _g,

$1-p_{\mathrm{fa}}={\∫}_0^{T_g}{f}_{G^0}\left(z\right)\mathrm{dz}---\left(4\right)$

In formula, represent G ⁰the probability density function of distribution range model:

$f_{G^0}\left(z\right)=\frac{{2n}^n\mathrm{\Γ}(n-\mathrm{\α}){\mathrm{\γ}}^{-\mathrm{\α}}{z}^{2n-1}}{\mathrm{\Γ}\left(n\right)\mathrm{\Γ}(-\mathrm{\α}){(\mathrm{\γ}+{\mathrm{nz}}^2)}^{n-\mathrm{\α}}}---\left(5\right)$

In formula, alpha, gamma is respectively shape and scale parameter ,-α > 0, γ > 0; Work as n=1, when haplopia data, the analytic solution of formula (4) are:

For looking data, because formula (4) is without analytic solution, adopt dichotomy to determine more

Step 14: according to the criterion shown in formula (7), in original image to each test pixel I _tadjudicate,

Wherein H ₁and H ₀represent respectively object pixel hypothesis and background pixel hypothesis, thus the target index matrix M shown in production (8):

$M(x,y)=\left\{\begin{array}{cc}1& ,H_1\\ 0& ,H_0\end{array}\right.---\left(8\right)$

In formula, (x, y) represents the coordinate position of pixel;

(2) adopt the G based on integral image ⁰distribution statistics parameter method for quick estimating, wherein, after trying to achieve the 2-rank, 4-rank integral image of original image, can try to achieve statistical parameter by simple operation;

Wherein, the G of described employing based on integral image ⁰distribution statistics parameter method for quick estimating, its concrete steps are as follows:

Step 21: by repeatedly carrying out following two groups of computings, the m-rank image i of an original image i of scanning (x, y) ^m(x, y), generates its m-rank integral image ii ^m(x, y),

s(x,y)＝s(x,y-1)+i ^m(x,y) (9)

ii ^m(x,y)＝ii ^m(x-1,y)+s(x,y) (10)

In formula, (x, y) represents the coordinate position of current pixel point; wherein i _k()=i (), k=1 ..., m, " dot product " of Π presentation video; S (x, y) represents i ^m(x, y) follows integration, wherein s (x ,-1)=0, ii ^m(1, y)=0; The needs of estimating according to statistical parameter, get m=2,4;

Step 22: the m-rank integral image based on original image, calculate fast the m-rank sample moment of reference unit in any two-parameter moving window,

$\begin{array}{c}{\mathrm{\μ}}^m=\frac{1}{N_s}\underset{t=1}{\overset{N_s}{\mathrm{\Σ}}}{\left|z\left(t\right)\right|}^m\\ =\frac{1}{N_s}\left\{\right[({\mathrm{ii}}_1^m+{\mathrm{ii}}_4^m)-({\mathrm{ii}}_2^m+{\mathrm{ii}}_3^m)]-[({\mathrm{ii}}_{1\′}^m+{\mathrm{ii}}_{4\′}^m)-({\mathrm{ii}}_{2\′}^m+{\mathrm{ii}}_{3\′}^m)\left]\right\}\end{array}----\left(11\right)$

In formula, for the pixel value of p place, summit m-rank integral image, p=1,2,3,4,1', 2', 3', 4'; corresponding to pixel value sum in whole moving window, for pixel value sum in protecting window;

Step 23: by 2-rank, 4-rank sample moment substitution formula (1) and (2), can try to achieve G by simple operation ⁰the statistical parameter of model;

(3), by the quick A CCA-CFAR algorithm based on integral image, realize SAR image object and detect;

Wherein, described pass through the quick A CCA-CFAR algorithm based on integral image, realize SAR image object and detect, its specific implementation step is as follows:

Step 301: adopt based on G ⁰the global threshold CFAR algorithm distributing carries out target pre-segmentation, generates target index matrix; For the ease of follow-up computing, to this index matrix negate, obtain background pixel index matrix, wherein background dot is 1, impact point is 0, for marking all candidate background pixels;

Step 302: by by background pixel index matrix and original image dot product, remove the interference of candidate target pixel, obtain background clutter image, with the reservation background pixel value of original image equidimension, the image of target pixel value zero setting;

Step 303: the m-rank integral image that calculates background clutter image;

Step 304: invariable false alerting and moving window parameter are set, and wherein the size of protecting window should be greater than the size of target to be detected, sliding window size should participate in statistical parameter estimation to ensure abundant background pixel fully greatly;

Step 305: utilize moving window by picture element scan original image, and by m-rank integral image and background pixel index matrix, calculate the m-rank sample moment of background pixel;

Step 306: utilize the m-rank sample moment of background pixel, adopt MoM method to estimate G ⁰the statistical parameter distributing;

Step 307: utilize statistical parameter to calculate CA-CFAR local threshold;

Step 308: by the size of compare test unit and local threshold, judge that whether this unit is candidate target pixel, is to be 1, otherwise is 0;

Step 309: judge whether to continue scanning, if do not scanned entire image, skip to step 305, scan next test cell; Otherwise, carry out next step;

Step 310: remove false-alarm by subsequent treatment, merge object pixel region, thereby obtain final target detection result; Subsequent treatment described in step 310 is counting wave filter, morphology processing or target area cluster.