CN105354824B

CN105354824B - DP-CFAR detection method based on extracted region

Info

Publication number: CN105354824B
Application number: CN201510641963.4A
Authority: CN
Inventors: 杜兰; 代慧; 王兆成; 肖金国
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2015-09-30
Filing date: 2015-09-30
Publication date: 2018-03-06
Anticipated expiration: 2035-09-30
Also published as: CN105354824A

Abstract

The invention discloses a kind of DP-CFAR detection method based on extracted region, mainly solves the problems, such as that detection speed is slow in existing SAR image target detection technique and parameter setting causes target missing inspection.Implementation step is：Positive and negative sample set is extracted to the training image with target label, trains a template w based on specification Gradient Features with linear classifier SVM, and effective set of dimensions is selected based on initial picture size to the positive sample collection of extraction；Template w and effective set of dimensions are then based on to the region under test image extraction effective dimensions；The region of extraction is detected with DP-CFAR again to obtain candidate region, and NMS is suppressed with non-maximum to candidate region, removes a large amount of overlapping regions, last remaining region is final testing result.The advantages of present invention detects compared to traditional two-parameter CFAR has detection speed fast and detection probability is high, suitable for the quick detection of SAR image target.

Description

DP-CFAR detection method based on extracted region

Technical field

The invention belongs to Radar Technology field, more particularly to a kind of CFAR detection method, available in synthetic aperture thunder Target is quickly and efficiently detected up in SAR image.

Background technology

Radar imaging technology is to grow up the 1950s, the hair advanced by leaps and bounds in 60 years afterwards Exhibition, at present, military affairs, agricultural, geology, ocean, disaster, paint survey etc. all many-sides be widely used.

Synthetic aperture radar SAR is a kind of active sensor perceived using microwave, it is with infrared, optics etc. other Sensor is compared, and SAR imagings are not limited by conditions such as illumination, weather, can carry out round-the-clock, whole day to target interested When observation, SAR image automatic target detection is by more and more extensive concern.

The tertiary treatment flow that U.S.'s Lincoln laboratory proposes is usually taken in SAR automatic target detection ATR methods.The flow Attention mechanism is layered using one kind, its implementation process is：First, detection process is carried out to view picture SAR image, removed bright in image Aobvious is not mesh target area, obtains potential target region；Then, target discriminating processing is carried out to potential target area, to pick Except natural clutter false-alarm therein, or reject region substantially bigger than target or small；Detection and discriminating stage by target, Obtain target region of interest ROI；Finally, then to target ROI Classification and Identification is carried out., it is necessary to handle in this treatment mechanism Data volume gradually reducing, can thus improve the efficiency of target identification system.

SAR image target detection is the first step in automatic target detection ATR flow scheme designs, and its importance is self-evident. How fast and effectively to detect potential target region is also a big study hotspot of the application of SAR image interpretation in recent years.

In existing SAR image object detection method, two-parameter CFAR detection algorithms are most widely used.It is two-parameter CFAR detection algorithms are a kind of traditional SAR image object detection methods, and the premise of this method application is the target in SAR image There is higher contrast with background clutter.Target window, protecting window and background are provided with two-parameter CFAR detection algorithms This 3 windows of window.Wherein, target window be possible the window containing object pixel, protecting window is to prevent target picture Element is mixed into the window set in background clutter, and backdrop window is the window containing background clutter.Two-parameter CFAR is based on the back of the body The statistical distribution pattern of scape clutter is the hypothesis of Gaussian Profile.By sliding window, to each pixel progress time in SAR image Go through.During each sliding window, by calculating average and the variance of all pixels in backdrop window come miscellaneous to background Ripple carries out parameter Estimation and determines a threshold value with this, if the pixel in target window is taken as mesh more than this threshold value Pixel is marked, it is clutter pixel to be otherwise considered as it.Because the algorithm needs that each pixel in SAR image is repeated to do identical Processing, it is long to cause this method detection time, while this method will set target according to the prior information of SAR image target Window, protecting window, backdrop window, for the too big target of size difference, parameter setting is unreasonable to cause background clutter to be joined Number estimation is inaccurate, so as to cause the missing inspection of target, while at a distance of closer target, as clustering distance it is unreasonable and It is considered as a target, appears in same detection zone, difficulty is caused to the subsequent treatment after detection.

The content of the invention

It is an object of the invention to for above-mentioned the deficiencies in the prior art, propose a kind of two-parameter perseverance based on extracted region False-alarm detection method, to reduce the detection time of target, improve the accuracy of detection of target.

To achieve the above object, technical scheme comprises the following steps：

(1) in markd training set Tr, positive sample collection P and negative sample collection N is extracted；

(2) the positive negative sample in obtained positive and negative sample set is down sampled to fixed dimension 8 × 8, to every after down-sampling Individual sample extraction specification Gradient Features g ', specification Gradient Features collection G, instruction are formed with the specification Gradient Features of all down-sampling samples Practice linear classifier SVM, obtain the template W of one 8 × 8；

(3) build effective set of dimensions and close AS：

(3a) initial 36 different picture size, form set S={ (W₁×H₁),...(W_l×H_l)...,(W₃₆× H₃₆), wherein, W_l,H_lIt is the wide and high of l-th picture size respectively, 1≤l≤36 and l is integer；These sizes using radix as 2, power increases successively from min T=3 to max T=8, i.e. W_l,H_l∈{8,16,32,64,128,256}；

The size set S that (3b) initializes according to the positive sample collection P and step (3a) obtained in step (1), obtain effectively Size set1≤As_iAs in≤36 corresponding set S_iIt is individual Element, 1≤i≤ns and i are integer, and ns represents the number of effective dimensions；

(4) regions of the test image J under effective dimensions collection is extracted：

The size of (4a) in effective dimensions set AS, to test image J down-samplings, obtain adopting under under different sizes Master drawing, and the standardization Gradient Features figure of down-sampling figure is extracted, form specification Gradient Features atlas { F₁,...F_i...,F_ns, its Middle F_iFor the Gradient Features figure that standardizes under i-th of effective dimensions, 1≤i≤ns and i is integer；

(4b) is with the template W obtained in step (2) to specification Gradient Features atlas { F₁,...F_i...,F_nsIn it is each Characteristic pattern carries out sliding window, obtains shot chart { s₁,...s_i...,s_ns, to each width shot chart, suppress NMS choosings with non-maximum Go out K local maximum, K region is extracted on test image J according to the position of local maximum, forms regional ensemble R_i, Same operation is done to ns size, finally gives ns × K region, forms set of regions R；

(5) the set of regions R obtained to step (4b), each region is considered as an entirety, with the region 80% most The average M of strong pixel value represents the region, estimates background clutter with the pixel value outside the region in the hollow frame of 3 pixel wides AverageAnd standard deviationAnd according to the false-alarm probability Pr of setting, the local detection threshold value Th in each region is obtained, region Average M elects candidate region as in the region more than local detection threshold value Th, obtains candidate region collection R '；

(6) NMS is suppressed with non-maximum to the candidate region collection R ' that step (5) obtains and removes a large amount of overlapping regions, remained Under region form regional ensemble R_d', as testing result.

The present invention compared with prior art, has advantages below：

1. detection speed is fast

Existing two-parameter CAFR detection methods detection when, due to will to each pixel sliding window in image, and The average and standard deviation of the model parameter, i.e. clutter of background clutter in estimation reference window, parameter Estimation are needed in each sliding window The mainly computing of addition, multiplication and division, so two-parameter CAFR algorithm complex is only relevant with the size of image.The present invention is A small amount of region is first obtained, DP-CFAR detection then is carried out to the region of extraction.DP-CFAR inspection is done to region It is surveyed unlike two-parameter CFAR detections using obtained region as integrally, rather than each pixel value is slided Window, its complexity are mainly relevant with the areal of extraction.

The extracted region of the present invention is write with C++ codes, while make use of the two-value of Gradient Features approximate, and then utilizes The bit manipulation of computer-internal, accelerates arithmetic speed, avoids the drawbacks of two-parameter CFAR is handled all pixels point, can be with Detection task is completed in the shorter time, accelerates detection speed.

2. the different target of image mesoscale can be handled, the accuracy of detection of target is improved

, it is necessary to be set using the prior information of target in image when existing two-parameter CFAR detection methods detect to image Parameter is put, if target size difference is too big in image, parameter setting is unreasonable, may result in the missing inspection of target, while also can It is a target to cause closely located target to be tested, and appears in a detection zone.And the present invention is the area based on extraction Domain is detected, and each region is considered as into an entirety, avoids in two-parameter CFAR that parameter setting is unreasonable to bring problem.

The present invention is described in further detail below in conjunction with drawings and examples.

Brief description of the drawings

Fig. 1 is the implementation process figure of the present invention；

Fig. 2 is the markd training set of band used in present invention experiment；

In Fig. 3 is the test set that uses in present invention experiment, and wherein Fig. 3 (a) is the test image containing 4 targets, Fig. 3 (b) it is the test image containing 7 targets, Fig. 3 (c) is the test image containing 3 targets；

Fig. 4 is the testing result obtained with the present invention to the test set in Fig. 3, and wherein Fig. 4 (a) is the detection to Fig. 3 (a) As a result, Fig. 4 (b) is the testing result to Fig. 3 (b), and Fig. 4 (c) is the testing result to Fig. 3 (c)；

It with traditional two-parameter CFAR under unified parameters is pair to the testing result of test set, wherein Fig. 5 (a) that Fig. 5, which is, Fig. 3 (a) testing result, Fig. 5 (b) are the testing results to Fig. 3 (b), and Fig. 5 (c) is the testing result to Fig. 3 (c)；

Embodiment

Reference picture 1, step is as follows for of the invention realizing：

Step 1, positive and negative sample set is extracted from markd training set.

(1.1) training set is inputted：

If training set Tr={ (I₁,B₁),...(I_j,B_j)...,(I_m,B_m),

Wherein, m is the number of training data, I_jIt is the jth width training image in training set,

B_j={ b₁ ^j,...b_k ^j...,b_tj ^jIt is that target target frame set is gazed in the acceptance of the bid of jth width training image,

Tj is the number of target in jth width image；

b_k ^j=(l, t, r, b) is the target frame of k-th of label target in jth width training image, wherein, l represents target frame The abscissa of upper left point, t represent the ordinate of target frame upper left point, and r represents the abscissa of target frame lower-right most point, and b represents target The ordinate of frame lower-right most point；

(1.2) in training image I_jMiddle extraction positive sample：

(1.2a) is in the B_jIn k-th of target b_k ^j=(l, t, r, b) extracts one group of regional ensemble Rb_k ^j={ rb₁, ...rb_n...,rb_nb, wherein:rb_nN-th of region of=(l ', t ', r ', b ') expressions, the abscissa of l ' expressions region upper left point, The ordinate of t ' expressions region upper left point, the abscissa of r ' expressions region lower-right most point, the ordinate of b ' expressions region lower-right most point,

Nb represents the number in region, nb=(w_max-w_min+1)×(h_max-h_min+ 1),

L '=l；T '=t；R '=l+2^p；B '=t+2^q

Wherein,

Function ceil (a) expressions take the smallest positive integral not less than a, and max (a, d) represents to take the maximum in both, min (a, d) represents to take the minimum value in both, and int () represents to ask business's computing, and mod () represents modulo operation；

(1.2b) is from regional ensemble Rb_k ^jMiddle extraction target b_k ^jThe positive sample region Pb at=(l, t, r, b) places_k ^j：

Zoning set Rb_k ^j={ rb₁,...rb_n...,rb_nbIn each region and target b_k ^j=(l, t, r's, b) covers Lid rate ovl, coverage rate ovl is more than or equal to 0.5 region, as target b_k ^jThe positive sample region of=(l, t, r, b) places extraction Pb_k ^j, wherein, n-th of region rb_n=(l ', t ', r ', b ') and target b_k ^j=(l, t, r, b) coverage rate ovl calculation formula For：

In formula, size () represent region in number of pixels and；

(1.2c) is according to step (1.2a) and (1.2b) to jth width training image target frame set B_j={ b₁ ^j, ...b_k ^j...,b_tj ^jIn all Objective extraction positive sample regions, form to training image I_jThe positive sample collection X of extraction_j=Pb₁ ^j ∪...∪Pb_k ^j∪...∪Pb_tj ^j；

(1.3) in training image I_jMiddle extraction negative sample collection：

(1.3a) generates four number x at random with randomly generated test problems rand ()₁、y₁、x₂、y₂, form region Nr₁=(x₁′, y₁′,x′₂,y′₂), wherein：

x₁The abscissa of ' expression region upper left point,

y₁The ordinate of ' expression region upper left point,

x₂The abscissa of ' expression region lower-right most point,

y₂The ordinate of ' expression region lower-right most point,

In formula, W, H divide table to represent training image I_jIt is wide and high；

(1.3b) zoning Nr₁=(x₁′,y₁′,x′₂,y′₂) and jth width training image target frame set B_j= {b₁ ^j,...b_k ^j...,b_tj ^jIn all target frames coverage rate ovl, if coverage rate ovl minimum value be less than 0.5, region Nr₁=(x₁′,y₁′,x′₂,y′₂) it is negative sample；

(1.3c) iteration (1.3a) and (1.3b) be after totally 50 times, obtained negative sample composing training image I_jNegative sample This set Y_j；

(1.4) to training set Tr={ (I₁,B₁),...(I_j,B_j)...,(I_m,B_m) in other training image repeat steps (1.2) and (1.3), positive sample collection P=X is obtained₁∪...∪X_j∪...∪X_mWith negative sample collection N=Y₁∪...∪Y_j∪...∪ Y_m。

Step 2, with the specification Gradient Features collection training linear classifier SVM of positive and negative sample set, the mould of one 8 × 8 is obtained Plate W

(2.1) the specification Gradient Features g ' of a sample in positive and negative sample set is extracted：

The positive and negative sample set that (2.1a) obtains to step 1 is uniformly down sampled to fixed dimension 8 × 8, after down-sampling Each sample, with one-dimensional horizontal gradient template A=[- 1,0,1] and vertical gradient template A^T, the sample is obtained in level side To gradient map g_x=F*A and vertical direction gradient map g_y=F*A^T, form the gradient map g=of the sample | g_x|+|g_y|；Its In, the sample in positive and negative sample set after F expression down-samplings, T represents to seek transposition, and * represents to seek convolution；

(2.1b) standardizes to gradient map g, obtains the gradient map g that standardizes_N=min (g, 255)；

The standardization Gradient Features figure g that (2.1c) will be obtained_NColumn vector is pulled into by row, obtains specification Gradient Features g '；

(2.2) the specification Gradient Features of all down-sampling samples are extracted, with the specification Gradient Features of all down-sampling samples Form specification Gradient Features collection G；

(2.3) specification Gradient Features collection G training linear classifier SVM are used, obtain the template W of one 8 × 8.

Step 3, according to positive sample collection, effective dimensions collection is selected.

(3.1) initial pictures set of dimensions：

Initial 36 different picture size, form set S={ (W₁×H₁),...(W_l×H_l)...,(W₃₆×H₃₆), Wherein, W_l,H_lIt is the wide and high of l-th picture size respectively, 1≤l≤36 and l is integer；These sizes are using radix as 2, power Increase successively from min T=3 to max T=8, i.e. W_l,H_l∈{8,16,32,64,128,256}；

(3.2) positive sample collection P=X is calculated₁∪...∪X_j∪...∪X_mIn each positive sample size label：

If region r_nIt is to jth width training image I_jThe positive sample collection X of extraction_jIn k-th of target b_k ^j=(l, t, r, b) places N-th of region, wherein, l represents the abscissa of target frame upper left point, and t represents the ordinate of target frame upper left point, and r represents mesh The abscissa of frame lower-right most point is marked, b represents the ordinate of target frame lower-right most point,

Then the size label sl of positive sample calculation formula is as follows：

Sl=6 × (q-min T)+(p-min T)+1,1≤sl≤36 and sl are integer,

Wherein,

(3.3) according to the size label of positive sample, positive sample number under each size label is counted, positive sample number is big It is considered as effective dimensions label in the size label equal to 5, it is { As to form effective dimensions label set₁,...As_i...,As_ns, its In, As_iIt is the size label of i-th of effective dimensions, 1≤As_iAs in≤36 corresponding set S_iIndividual element, 1≤i≤ns and i is Integer, ns represent the number of effective dimensions；

(3.4) in initial pictures size set S={ (W₁×H₁),...(W_l×H_l)...,(W₃₆×H₃₆) in selection the As_iIndividual elementObtain i-th of effective dimensionsEffective dimensions collection is obtained by effective dimensions label set

Step 4, region of the test image under effective dimensions collection is extracted.

(4.1) according to effective dimensions collectionTo test chart As J progress down-samplings, the image after one group of down-sampling is obtained After being i-th of effective dimensions down-sampling Image, the size of the image is

Wherein,Represent i-th of down-sampled images Js_iWidth,

Represent i-th of down-sampled images Js_iHeight,

In formula, W ', H ' are the wide and high of test image J respectively, and function ceil (a) represents to take the smallest positive integral not less than a；

(4.2) to the image Js after down-sampling_iExtraction specification gradient map：F_i=min (| Js_i*A|+|Js_i*A^T|, 255), its In, A=[- 1,0,1], T represent transposition, and * represents to seek convolution；

(4.3) with template W to the gradient map { F that standardizes₁,...F_i...,F_nsIn each width characteristic pattern carry out sliding window, obtain To shot chart { s₁,...s_i...,s_ns, wherein s_i=W*F_i, * represent ask convolution, F_iRepresent the specification under i-th of effective dimensions Gradient map, s_iRepresent the shot chart under i-th of effective dimensions；

(4.4) shot chart s of the NMS algorithms in i-th of size is suppressed using non-maximum_iIn select the K of the shot chart Local maximum, location sets Ms is formed with the coordinate position of K local maximum of the shot chart_i：

(4.4a) is by shot chart s_iArranged from big to small by its score value, the shot chart s after being arranged_i', and arrange Component s_i' in all scoring positions mark and be；

(4.4b) will arrange shot chart s_iThe score position coordinate that ' middle position mark is true and score value is maximum is put into Location sets Ms_iIn, and be false by four neighbourhood signatures of the position and the position；

(4.4c) continues step (4.4b), until obtaining the coordinate position of K local maximum, forms location sets Ms_i ={ (u₁′,v₁′),...(u_j′,v_j′)...,(u_K′,y′_K), (u_j′,v_j') represent j-th of local maximum point in shot chart Position coordinates, 1≤j≤K and j is integer；

(4.5) according to location sets Ms_i, test image J is in i-th of size for extractionUnder K region, form Regional ensembleWhereinJ-th of region under i-th of size is represented,

v₁The abscissa of region upper left point is represented,

v₂The ordinate of region upper left point is represented,

v₃The abscissa of region lower-right most point is represented,

v₄The ordinate of region lower-right most point is represented,

In formula, function ceil (a) represents to take the smallest positive integral not less than a；

(4.6) to shot chart { s₁,...s_i...,s_nsIn shot chart under other sizes according to step (4.4) and (4.5) K region is extracted under each effective dimensions；

(4.7) regional ensemble that test image obtains under ns effective dimensions：

R=R₁∪...∪R_i∪...∪R_ns, wherein, R_iRepresent the regional ensemble extracted under i-th of effective dimensions.

Step 5, candidate region is obtained.

(5.1) to each region in regional ensemble R, this is represented with the average M of 80% most strong pixel value in the region Region；

(5.2) with the average of the calculated for pixel values background clutter outside each region in the hollow frame of 3 pixel widesWith Standard deviationCalculation formula is as follows：

In formula, Ω_cRepresent clutter region, N_cRepresent clutter area pixel number；

(5.3) the local detection threshold value Th in each region is calculated,Wherein K_CFARAccording to setting False-alarm probability Pr is calculated_,Calculation formula is：K_CFAR=Φ^-1(1-Pr), wherein Φ () are Standard Normal Distributions, Φ^-1() is the inverse function of Standard Normal Distribution；

(5.4) the average M in each region is compared with the local detection threshold value Th in the region, if Th ＞ M, retain The region, otherwise, the region is deleted, the region finally given forms candidate region collection R '.

Step 6, overlapping candidate region is removed, obtains testing result.

(6.1) collection R ' middle all areas in candidate region are arranged from big to small by its average M, the candidate regions after being arranged Domain collection R₁′；

(6.2) calculated permutations candidate region collection R₁' in first region and other regions coverage rate, from arrangement candidate Set of regions R₁' middle the region for removing coverage rate and being more than or equal to 0.01, and by the R₁' in first region be put into regional ensemble R_d′ In；

(6.3) renewal arrangement candidate region collection R₁'=R₁′-R_d', and in return to step (6.2), until arranging candidate region Collect R₁' for sky, the regional ensemble R finally given_d' it is testing result.

The effect of the present invention can be further illustrated by following experiment：

1. experiment condition

Test operation platform：MATLAB R2012a,Visual Studio 2012,Intel(R)Core(TM)i5-4590 CPU@3.30GHZ, Windows7 Ultimates.

Data used in experiment are the data of complete polarization and single polarization in RADARSAT-2 databases, the training set shown in Fig. 2 It is that interception comes out in RADARSAT-2 data with the test set shown in Fig. 3.Rectangle frame marks in training set and test set It is the target frame of target in image.Target in data set is all naval vessel, and the size on naval vessel differs, and list is used as using pixel number Position, wherein target full-size a width of 96, a height of 16, target minimum dimension a width of 12, a height of 8.

In experiment, the parameter setting of detection method：The number K=150, false alarm rate Pr=in region under each size 10^-2, initial pictures size set S={ (W₁×H₁),...(W_l×H_l)...,(W₃₆×H₃₆), wherein W_l,H_lIt is l-th respectively Wide and high, 1≤l≤36 and l is integer of picture size；These sizes are using radix as 2, and power is from min T=3 to max T=8 Increase successively, i.e. W_l,H_l∈{8,16,32,64,128,256}

In experiment, the parameter setting of two-parameter CFAR detection methods：False alarm rate Pr=10^-6, next sample number r=2, protection Half long mG=27 of window, half long mB=30 of background window.Minimum target region area S_min=4, maximum target region area S_max= 300, target maximum length Len=25, slice size Q=100.

2. experiment content：

Experiment 1, is first trained with the inventive method to the training set shown in Fig. 2, and then the test set in Fig. 3 is carried out Target detection, as a result as shown in Figure 4；

Experiment 2, target detection is carried out with traditional two-parameter CFAR detection methods to the test image in Fig. 3, two-parameter Master thesis of the concrete operations of CFAR detections with reference to Xian Electronics Science and Technology University Li Li in 2013《SAR target detections with Identification algorithm is studied and Software for Design》In chapter 2 " SAR target detection method research ", testing result as shown in figure 5, its In, Fig. 5 (a) is the testing result to Fig. 3 (a), and Fig. 5 (b) is the testing result to Fig. 3 (b), and Fig. 5 (c) is the inspection to Fig. 3 (c) Survey result.

The testing result such as table 1 of experiment 1 and experiment 2：

The inventive method of table 1 and two-parameter CFAR contrast to test image detection case

3. analysis of experimental results

As can be seen from Table 1, for testing SAR image data used, the present invention applies double ginsengs based on extracted region Number CFAR detection method has carried out target detection, shows the detection method of the present invention and has good performance.Test chart in table 1 During as being detected with traditional two-parameter CFAR, because parameter setting is unreasonable, there is the missing inspection of target, to Fig. 3 from table 1 (a) and Fig. 3 (b) test images testing result can be seen that target number more than missing inspection number with detection target area number Sum；

It can also be seen that the time used in the inventive method detection test image is about traditional two-parameter CFAR from table 1 The 1/3 of detection time, because the inventive method is in the training stage and to obtain the time used in the region of test image very short, therefore The present invention improves detection efficiency to a certain extent.

Comparison diagram 3 and testing result Fig. 5 are it can also be seen that the missing inspection of target all occur in three width test images, from Fig. 5 (a) and Fig. 5 (b) is it can be seen that have two targets to appear in a detection zone, mainly due to target at a distance of relatively near, parameter Set unreasonable caused.It can be seen that the present invention compared to traditional two-parameter CFAR detection methods for, can not only ensure compared with High verification and measurement ratio, while be not in situation of multiple targets in a detection zone.Comparison diagram 4 and Fig. 5 can be seen that this hair Bright SAR image object detection method has the advantages of detection probability is high, while testing result is region not of uniform size, can pin The target to differ to size.

To sum up, SAR image object detection method of the invention has the advantages of algorithm execution speed is fast, detection probability is high, It is a kind of quick and effective detection method, and it is adaptable to various sizes of target, before there is good application Scape.

Claims

1. the DP-CFAR detection method based on extracted region, comprises the following steps：

(2) the positive negative sample in obtained positive and negative sample set is down sampled to fixed dimension 8 × 8, to each sample after down-sampling This Extraction specificationization Gradient Features g ', the standardization of sample image is formed with the standardization Gradient Features figure of all down-sampling samples Gradient Features atlas G, training linear classifier SVM, obtain the template W of one 8 × 8；

(3) build effective set of dimensions and close AS：

(3a) initial 36 different picture size, form set S={ (W₁×H₁),...(W_l×H_l)...,(W₃₆×H₃₆), Wherein, W_l,H_lIt is the wide and high of l-th picture size respectively, 1≤l≤36 and l is integer；These sizes are using radix as 2, power Increase successively from minT=3 to maxT=8, i.e. W_l,H_l∈{8,16,32,64,128,256}；

The size set S that (3b) initializes according to the positive sample collection P and step (3a) obtained in step (1), obtains effective dimensions Set1≤As_iAs in≤36 corresponding set S_iIndividual member Element, 1≤i≤ns and i are integer, and ns represents the number of effective dimensions；

The size of (4a) in effective dimensions set AS, to test image J down-samplings, obtains the down-sampling under different sizes Figure, and the standardization Gradient Features figure of down-sampling figure is extracted, form the standardization Gradient Features atlas { F of test image₁, ...F_i...,F_ns, wherein F_iFor the Gradient Features figure that standardizes under i-th of effective dimensions；

(4b) is with the template W obtained in step (2) to the Gradient Features atlas { F that standardizes₁,...F_i...,F_nsIn each feature Figure carries out sliding window, obtains shot chart { s₁,...s_i...,s_ns, to each width shot chart, suppress NMS with non-maximum and select K Local maximum, K region is extracted on test image J according to the position of local maximum, forms regional ensemble R_i, to ns Size does same operation, finally gives ns × K region, forms set of regions R；

(5) the set of regions R obtained to step (4b), each region is considered as an entirety, with the region 80% most strong picture The average M of plain value represents the region, estimates background clutter average with the pixel value outside the region in the hollow frame of 3 pixel widesAnd standard deviationAnd according to the false-alarm probability Pr of setting, the local detection threshold value Th in each region is obtained, regional average value M Candidate region is elected in region more than local detection threshold value Th as, obtains candidate region collection R '；

(6) NMS is suppressed with non-maximum to the candidate region collection R ' that step (5) obtains and removes a large amount of overlapping regions, it is remaining Region forms regional ensemble R_d', as testing result.

2. according to the method for claim 1, wherein positive and negative sample set is extracted in training set in the step (1), by such as Lower step is carried out：

(1a) defines training set：

If training set Tr={ (I₁,B₁),...(I_j,B_j)...,(I_m,B_m), wherein m is the number of training data, I_jIt is training set In jth width training image, B_j={ b₁ ^j,...b_k ^j...,b_tj ^jIt is that target target frame set is gazed in the acceptance of the bid of jth width training image, Tj is the number of target in jth width training image；b_k ^j=(l, t, r, b) is the mesh of k-th of label target in jth width training image Frame is marked, wherein, l represents the abscissa of target frame upper left point, and t represents the ordinate of target frame upper left point, and r represents target frame bottom right The abscissa of point, b represent the ordinate of target frame lower-right most point；

(1b) is in training image I_jMiddle extraction positive sample：

(1b.1) is to B_jIn k-th of target b_k ^j=(l, t, r, b) extracts one group of region Rb_k ^j={ rb₁,...rb_n...,rb_nb, Wherein, nb represents the number in region, rb_n=(l ', t ', r ', b ') represents n-th of region, the horizontal seat of l ' expressions region upper left point Mark, the ordinate of t ' expressions region upper left point, the abscissa of r ' expressions region lower-right most point, the vertical seat of b ' expressions region lower-right most point Mark；

(1b.2) is from regional ensemble Rb_k ^j={ rb₁,...rb_n...,rb_nbIn extraction target b_k ^jThe positive sample at=(l, t, r, b) places Region Pb_k ^j；

(1b.3) is according to (1b.1) and (1b.2) to B_j={ b₁ ^j,...b_k ^j...,b_tj ^jIn all Objective extraction positive sample areas Domain, form to training image I_jThe positive sample collection X of extraction_j=Pb₁ ^j∪...∪Pb_k ^j∪...∪Pb_tj ^j；

(1c) is in training image I_jMiddle extraction negative sample collection：

(1c.1) generates four numbers at random, forms region Nr₁=(x₁′,y₁′,x₂′,y₂'), wherein, x₁' represent region upper left point Abscissa, y₁The ordinate of ' expression region upper left point, x₂The abscissa of ' expression region lower-right most point, y₂' represent region lower-right most point Ordinate；

(1c.2) zoning Nr₁=(x₁′,y₁′,x₂′,y₂') and training image I_jIn all target B_j={ b₁ ^j,...b_k ^j..., b_tj ^jCoverage rate ovl, if coverage rate ovl minimum value be less than 0.5, region Nr₁=(x₁′,y₁′,x₂′,y₂') it is negative Sample；

(1c.3) iteration (1c.1) and (1c.2) be after totally 50 times, obtained negative sample composing training image I_jNegative sample collection Close Y_j；

(1d) is to training set Tr={ (I₁,B₁),...(I_j,B_j)...,(I_m,B_m) in other training image repeat steps (1b) and (1c), obtain positive sample collection P=X₁∪...∪X_j∪...∪X_mWith negative sample collection N=Y₁∪...∪Y_j∪...∪Y_m。

3. according to the method for claim 2, the wherein areal nb in step (1b.1), it is calculated as follows：

Nb=(w_max-w_min+1)×(h_max-h_min+1)

Wherein,

Function ceil (a) expressions take the smallest positive integral not less than a, and max (a, d) represents to take the maximum in both, min (a, d) Represent to take the minimum value in both.

According to the method for claim 2,4. wherein from regional ensemble Rb in step (1b.2)_k ^jMiddle extraction target b_k ^jPlace is just Sample areas Pb_k ^j, it is zoning set Rb_k ^j={ rb₁,...rb_n...,rb_nbIn each region and target b_k ^j=(l, t, R, b) coverage rate ovl, by coverage rate ovl be more than or equal to 0.5 region, as target b_k ^jExtract just at=(l, t, r, b) places Sample areas Pb_k ^j,

Wherein, n-th of region rb_n=(l ', t ', r ', b ') and target b_k ^j=(l, t, r, b) coverage rate ovl calculation formula For：

Size () represent region in number of pixels and.

5. according to the method for claim 1, wherein to each sample extraction specification after down-sampling in the step (2)Change Gradient Features g ', carry out as follows：

(2a) uses one-dimensional horizontal gradient template A=[- 1,0,1] and vertical to each sample after down-sampling respectively Gradient template A^TTo its gradient map g in the horizontal direction of the sample extraction_x=F*A and vertical direction gradient map g_y=F*A^T, structure Into the gradient map g=of the sample | g_x|+|g_y|, wherein, the sample in positive and negative sample set after F expression down-samplings, T tables Show and seek transposition, * represents to seek convolution；

(2b) standardizes to gradient map g, obtains the Gradient Features figure g that standardizes_N=min (g, 255)；

The standardization Gradient Features figure g that (2c) will be obtained_NColumn vector is pulled into by row, obtains specification Gradient Features g '.

6. according to the method for claim 1, wherein extracting the standardization Gradient Features of down-sampling figure in the step (4a) Figure, is carried out as follows：

(4a.1) is according to effective dimensions collectionTest image J is entered Row down-sampling, obtain the image { Js after one group of down-sampling₁,...Js_i...,Js_ns, Js_iAfter being i-th of effective dimensions down-sampling Image, the size of the image isCalculation formula is as follows：

<mrow> <msubsup> <mi>W</mi> <mrow> <msub> <mi>As</mi> <mi>i</mi> </msub> </mrow> <mo>&prime;</mo> </msubsup> <mo>=</mo> <mi>c</mi> <mi>e</mi> <mi>i</mi> <mi>l</mi> <mrow> <mo>(</mo> <mn>8</mn> <mo>&times;</mo> <msup> <mi>W</mi> <mo>&prime;</mo> </msup> <mo>/</mo> <msub> <mi>W</mi> <mrow> <msub> <mi>As</mi> <mi>i</mi> </msub> </mrow> </msub> <mo>)</mo> </mrow> <mo>,</mo> <msubsup> <mi>H</mi> <mrow> <msub> <mi>As</mi> <mi>i</mi> </msub> </mrow> <mo>&prime;</mo> </msubsup> <mo>=</mo> <mi>c</mi> <mi>e</mi> <mi>i</mi> <mi>l</mi> <mrow> <mo>(</mo> <mn>8</mn> <mo>&times;</mo> <msup> <mi>H</mi> <mo>&prime;</mo> </msup> <mo>/</mo> <msub> <mi>H</mi> <mrow> <msub> <mi>As</mi> <mi>i</mi> </msub> </mrow> </msub> <mo>)</mo> </mrow> </mrow>

Wherein, W ', H ' are the wide and high of test image J respectively, and function ceil (a) represents to take the smallest positive integral not less than a；

(4a.2) is to the image Js after down-sampling_iExtraction specification Gradient Features figure：F_i=min (| Js_i*A|+|Js_i*A^T|, 255), wherein, A=[- 1,0,1], T represent transposition, and * represents to seek convolution.

7. according to the method for claim 1, wherein extraction test image is under an effective dimensions in the step (4b) Region, carry out as follows：

(4b.1) is with template W to the Gradient Features atlas { F that standardizes₁,...F_i...,F_nsIn each width characteristic pattern carry out sliding window, Obtain shot chart { s₁,...s_i...,s_ns, wherein s_i=W*F_i, * represent seek convolution；

(4b.2) suppresses NMS to the shot chart s under i-th of size using non-maximum_iHandled, select in the shot chart K Local maximum, location sets Ms is formed with its position coordinates_i={ (u₁′,v₁′),...(u_j′,v_j′)...,(u_K′,v_K'), (u_j′,v_j') represent position coordinates of j-th of local maximum point in shot chart, 1≤j≤K and j is integer；

(4b.3) is according to location sets Ms_i, test image J is in i-th of size for extractionUnder K region, form area Gather in domainWhereinJ-th of region under i-th of size is represented,

v₁The abscissa of region upper left point is represented,

v₂The ordinate of region upper left point is represented,

v₃The abscissa of region lower-right most point is represented,

v₄The ordinate of region lower-right most point is represented,

Function ceil (a) represents to take the smallest positive integral not less than a.

8. according to the method for claim 1, wherein the step (6) suppresses NMS according to non-maximum removes candidate region Collect region overlapping in R ', carry out as follows：

(6a) is arranged the middle all areas of candidate region collection R ' by its average M from big to small, the candidate region collection after being arranged R₁′；

(6b) calculated permutations candidate region collection R₁' in first region and other regions coverage rate, from arrangement candidate region collection R₁' middle the region for removing coverage rate and being more than or equal to 0.01, and by the R₁' in first region be put into regional ensemble R_d' in；

(6c) renewal arrangement candidate region collection R₁'=R₁′-R_d', and in return to step (6b), until arranging candidate region collection R₁' be Sky, the regional ensemble R finally given_d' it is testing result.