CN105956592B - A kind of Aircraft Targets detection method based on saliency and SVM - Google Patents

Abstract

Translated fromChinese

本发明涉及一种基于图像显著性与SVM的飞机目标检测方法，显著性区域检测中有两种视觉注意方式，分别是：自底向上(数据驱动)的注意方式、自顶向下(任务驱动)的注意方式。本发明先提取训练样本的方向梯度直方图(HOG)特征，训练支持向量机(SVM)分类器，然后使用一种基于剩余谱理论的自底向上的视觉注意模型进行显著性区域检测，快速提取可能存在目标的候选区域，然后提取候选区域的HOG特征，再使用训练好的SVM分类器对候选区域进行分类，完成目标检测。

The invention relates to an aircraft target detection method based on image saliency and SVM. There are two visual attention modes in saliency area detection, namely: a bottom-up (data-driven) attention mode and a top-down (task-driven) attention mode. ) of attention. The invention first extracts the histogram of orientation gradient (HOG) feature of the training sample, trains the support vector machine (SVM) classifier, and then uses a bottom-up visual attention model based on the residual spectrum theory to detect the saliency region, and quickly extracts There may be candidate regions of the target, and then extract the HOG features of the candidate regions, and then use the trained SVM classifier to classify the candidate regions to complete the target detection.

Description

A kind of Aircraft Targets detection method based on saliency and SVM

Technical field

The invention belongs to Computer Image Processing, are related to Aircraft Targets detection method, and in particular to a kind of aobvious based on imageThe Aircraft Targets detection method of work property and SVM.

Background technique

Visible images target detection is important one of the branch of object detection field, is had in military field particularly importantUsing.Currently, experts and scholars both domestic and external have done a large amount of research work in terms of Airplane detection.Than there is base earlierIn the airplane detection method of image angle point and edge feature, aircraft is judged by face shaping that angle point and edge are surroundedPosition.In addition, there is a kind of circle filtering for Airplane detection in the prior art in the unique shape due to aircraft in the pictureDevice detection method；But this method is easy to be limited by aircraft size, to the inappropriate aircraft region of scale using circle filteringThe detection effect that device can not reach.And the airplane detection method based on saliency, then mainly utilize the vision of human eyeSystem features tentatively obtain the aircraft suspicious region in image, and the inspection realized to aircraft position is then combined with target signatureIt surveys.

In the method based on statistical learning, it is seen that the target detection in light image is usually identified as one two classification and asksTopic, i.e., target to be detected or be target or be not target.With the methods of machine learning and feature extraction research it is continuousIt deeply and is constantly applied in type target detection, existing machine learning method examines target with feature extracting methodThe accuracy of survey is greatly improved, but still cannot efficiently and accurately extract the candidate region in image comprising target.

Summary of the invention

Technical problems to be solved

In order to avoid the shortcomings of the prior art, the present invention proposes a kind of aircraft mesh based on saliency and SVMDetection method is marked, it is not high to solve detection accuracy by object detection field is applied to based on the image segmentation that salient region detectsThe problems such as.

Technical solution

A kind of Aircraft Targets detection method based on saliency and SVM, it is characterised in that steps are as follows:

Step 1, data preparation: visible light Airport Images are chosen from publicly available all kinds of images as training sampleAnd target atlas to be detected, and grayscale image is converted to, the size of the external square of maximum of aircraft is from minimum value in figure: a₁×a₁,It is a to maximum value₂×a₂；Selected part image will be used as positive sample after the cutting of aircraft region in image as training sample,Negative sample is used as after his region cutting；Remaining parts of images is as test sample；The sample of the training is scaled a₃×a₃The size of size；

Step 2: extracting the HOG feature of positive negative training sample respectively；

Step 3: by the HOG feature of each the positive negative sample extracted and its class label, positive sample 1, negative sampleA vector, training SVM classifier are combined into for -1；

Step 4: the notable figure of image in test sample is obtained using the conspicuousness detection method based on remaining spectral theory；

Step 5: the connected domain in notable figure is extracted by area threshold method, generates the candidate region of Aircraft Targets,Process is as follows:

Step 5a: the mean value of notable figure is calculated firstAnd variance

Wherein, w indicates that the width of image, h indicate that the height of image, Sal (i, j) indicate the pixel of the i-th row jth column in notable figureValue；

Step 5b: the mean value of notable figure is utilizedAnd varianceThreshold value T is calculated, notable figure is divided with threshold value T；The thresholdValue T:

Coefficient k is in order to which tension metrics are poorAnd mean valueWeight in threshold value T value

Step 5c, connected domain filtering is carried out to the notable figure after thresholding: is n with size₁×n₁Rectangular window, with the i-th rowA neighborhood is intercepted centered on jth column pixel (i, j)；The area that the region is defined with the number of pixels in certain region finds out placeIn the area of the marking area in the contiguous range；If the area is greater than certain threshold value, just retains the region, otherwise do not protectIt stays；Using all regions remained as the candidate region of Aircraft Targets；

Step 6: Aircraft Targets are extracted from the candidate region of Aircraft Targets, process is as follows:

Step 6a, response matrix is constructed:

A two-dimensional response matrix M is constructed, the value of each element is a two-dimensional array M in matrix_i,j(s, r) itsMiddle i ∈ [1, w], j ∈ [1, h], s are used for record window size, and r is used to record SVM caused by the window there are Aircraft TargetsResponse；When initialization, which is initialized as a full 0 matrix identical with image size；

Using the local pixel maximum of any candidate region as the center of the candidate region, referred to as: seed point；

Step 6b, secondary windows method rejects invalid seed point, retains most probable aircraft region, process is as follows:

(A) determine first time image block window: window is a rectangular window centered on seed point, the size of windowFor h₁×h₁, in which:

Wherein, int () function representation round.Original image is intercepted with the window, the figure that will be truncated toAs block is scaled a₃×a₃Size extracts the HOG feature of image block, then HOG feature is updated in SVM classifier, calculatesSVM classifier response r₁；

(B) determine second of image block window: window is a rectangular window centered on seed point, the size of windowFor h₂×h₂, in which:

h₂=2 (h₁-1)+1

Original image is intercepted with the window, the image block being truncated to is scaled a₃×a₃Size extracts image blockThen HOG feature is updated in SVM classifier by HOG feature, calculate SVM classifier response r₂；

(C) judge whether to retain the region:

If the classifier response being calculated twice is respectively less than 0, then it is assumed that the corresponding aircraft candidate regions of the seed pointDomain is invalid, therefore rejects the invalid seed point；

Otherwise it is assumed that the region that window includes there are aircraft or contains most of fuselage of aircraft, by two secondary responseIn the larger value assignment into response matrix on the corresponding coordinate position of seed point, the referred to as seed point of response matrix, and recordingThe lower corresponding window size of the response；Specifically, if changing coordinates are as follows: (u, v), the SVM classifier response that needs are recorded,It is assigned to the r value at the seed point of response matrix in the two-dimensional array of element M (u, v)；Corresponding interception window when the response will be generatedSquare side length value in mouth size, is assigned to the s value at the seed point of response matrix in the two-dimensional array of element M (u, v)；

Step 6c, optimize response matrix, determine aircraft region:

(1) optimize the window size recorded in each seed point:

To the seed point of each response matrix, transformed edge of window long value h is first calculated according to the following formula₃:

Centered on the coordinate at the seed point of response matrix, with h₃×h₃Interception window is established to original image for window sizeAs being intercepted, the image block being truncated to is scaled a₃×a₃Size extracts the HOG feature of image block, then by HOG featureIt is updated in SVM classifier, calculates SVM classifier response r₃；

Transformed edge of window long value h is calculated further according to following formula₄:

Centered on the coordinate at the seed point of response matrix, with h₄×h₄Interception window is established to original image for window sizeAs being intercepted, the image block being truncated to is scaled a₃×a₃Size extracts the HOG feature of image block, then by HOG featureIt is updated in SVM classifier, calculates SVM classifier response r₄；

Find out r, r₃,r₄In maximum value, replace at the seed point of response matrix original r value in the two-dimensional array of element,With original s value in the two-dimensional array of element at the seed point of the corresponding edge of window long value replacement response matrix of the new r value；

It determines aircraft region, obtains final target detection result；

(2) according to the response matrix after optimization, the coordinate recorded using at the seed point of response matrix is as winged in original imageThe center of square area where machine target, pros where the s value recorded using at the seed point as Aircraft Targets in original imageThe side length in shape region can determine each aircraft region, obtain final Aircraft Targets testing result.

Beneficial effect

A kind of Aircraft Targets detection method based on saliency and SVM proposed by the present invention, salient region detectionIn there are two types of vision attention mode, be respectively: the attention mode of bottom-up (data-driven), top-down (task-driven)Pay attention to mode.The present invention first extracts histograms of oriented gradients (HOG) feature of training sample, Training Support Vector Machines (SVM) pointClass device, then the bottom-up visual attention model using a kind of based on remaining spectral theory carries out salient region detection, fastlyThere may be the candidate regions of target for speed extraction, then extract the HOG feature of candidate region, reuse trained svm classifierDevice classifies to candidate region, completes target detection.

The present invention will dexterously be tied based on the detection of the saliency of remaining spectral theory and the learning method of support vector machinesAltogether, the selection for improving target area improves the robustness of detection accuracy and detection, and model is simple, and execution efficiency is high.

Detailed description of the invention

Fig. 1: the training flow chart of the detection network based on SVM

Fig. 2: the visible images Airplane detection flow chart based on saliency and SVM

Specific embodiment

Now in conjunction with embodiment, attached drawing, the invention will be further described:

Steps are as follows for embodiment:

Step 1: data preparation.

Visible light Airport Images are intercepted from GOOGLE-EARTH as training sample and target to be detected, are uniformly converted toGrayscale image, the size of the external square of maximum of aircraft is from minimum value in figure: a₁×a₁, until maximum value is a₂×a₂Differ.At thisIn embodiment, a₁=20, a₂=80.From the image after these interceptions, positive negative sample (each 2000) of the acquisition for training,Manually mark its aircraft region.Positive sample is the aircraft of various postures, different model, and negative sample is to build present in airportThe intersection etc. of road in mark, airplane tail group, airport in object, airport.In training, sample standard deviation used is uniformly contractedIt puts as a₃×a₃The size of size does not need to cut for every width test image.In the present embodiment, a₃=64.

Step 2: extracting the HOG feature of positive negative training sample.

First using Gamma correction method to after scaling having a size of a₃×a₃Image normalization.Then by image uniformIt is divided into b₁×b₁A " cell " (cell), the size of each cell are c₁×c₁A pixel, in which: c₁=a₃/b₁.For justIn calculating, usually by a₃,b₁,c₁Value be adjusted to integer value, in the present embodiment, b₁=8, c₁=8.It finds out each in cellThe gradient direction of pixel projects in corresponding histogram according to the amplitude weighting of its gradient direction, a cell pairAnswer the vector of a N-dimensional (N=9 here extracts 9 dimensional features from each cell).Then take b₂×b₂A cell combinationAt a block (block), the interval between block and the initial position of block takes d₁A pixel, so that mutual is overlapped between block, blockSelection sequence are as follows: first from left to right, then from top to bottom.In the present embodiment, b₂=2, d₁=8, it may be assumed that horizontal direction and Vertical SquareTo there is 7 blocks respectively, so that image includes 49 blocks altogether.Each cell still takes N-dimensional feature, and the feature of block is by the straight of cellSquare figure feature vector is composed in series according to putting in order for cell；The HOG feature of whole image is then by each piece of HOG featureVector is composed in series.In the present embodiment, have 9 × 2 × 2 totally 36 dimensional feature, entire normalized image share 36 × 49 for each pieceTotally 1764 dimensional feature.

Step 3: by the HOG feature of each the positive negative sample extracted in previous step and its class label, (positive sample is1, negative sample is -1) it is combined into a vector, training SVM classifier.The kernel function of SVM is using radial base core letter in the present embodimentNumber.

Step 4: after training, obtaining the significant of test image using the conspicuousness detection method based on remaining spectral theoryFigure.

(1) width of gray level image I to be detected is set as w pixel, is highly h pixel, to the pixel of image

Set I (x) (wherein [0,255] I (x) ∈, x ∈ [1, w*h]) carries out Fourier transformation F (I (x)), and extracts figurePicture frequency

The phase property and amplitude characteristic in domain；

A (f)=R (F (I (x))) (1)

P (f)=S (F (I (x))) (2)

Wherein, A (f) indicates that the amplitude of frequency f, P (f) indicate the phase of frequency f, and R (F (I (x))) expression takes F (I (x))Amplitude, S (F (I (x))) expression take F (I (x)) phase.

(2) log of calculated amplitude composes L (f), and composes to log and carry out mean filter, the residual spectra R (f) for then asking log to compose:

L (f)=log (A (f)) (3)

R (f)=L (f)-h_n(f)*L(f) (4)

H in above formula_nIt (f) be size is n × n (in the present embodiment select 3 × 3 sizes) mean filter convolution kernel, definitionFor

(3) residual spectra is acquired to new image spectrum in conjunction with phase spectrum, Fourier inversion is carried out to the image spectrumNew image is acquired, and image smoothing is carried out using Gaussian convolution core to obtained image, can be obtained based on remaining spectral theoryImage saliency map S (x):

S (x)=g (x) * F^-1[exp(R(f)+P(f))]² (5)

In above formula, F^-1It (f) is Fourier inversion, g (x) indicates Gaussian convolution core.

Step 5: the connected domain in notable figure being extracted by area threshold method, the candidate regions of Aircraft Targets are generated with thisDomain.

(1) mean value of notable figure is calculated firstAnd variance

Wherein, w indicates that the width of image, h indicate that the height of image, Sal (i, j) indicate the pixel of the i-th row jth column in notable figureValue.

(2) mean value of notable figure is utilizedAnd varianceThreshold value T is calculated, notable figure is divided with threshold value T.Here threshold valueT:

Coefficient k is in order to which tension metrics are poorAnd mean valueWeight in threshold value T value, in the present embodiment:The such value of k can make threshold value T and standard deviationMeet certain inverse proportionality characteristics, the difference that debases the standard is excessiveWhen influence to threshold value, or can the influence appropriate that increase standard deviation when standard deviation is too small.

(3) connected domain filtering is carried out to the notable figure after thresholding.It is n with size₁×n₁Rectangular window, with the i-th row jthA neighborhood, in the present embodiment, n are intercepted centered on column pixel (i, j)₁=11.The area is defined with the number of pixels in certain regionThe area in domain can find out the area of the marking area in the contiguous range.If the area is greater than certain threshold value (this realityIt applies in example, 60) threshold value takes, just retain the region, otherwise do not retain.Using all regions remained as the time of Aircraft TargetsFavored area.

Step 6: extracting Aircraft Targets from the candidate region of Aircraft Targets.

(1) response matrix is constructed

A two-dimensional response matrix M is constructed, the value of each element is a two-dimensional array M in matrix_i,j(s, r) itsMiddle i ∈ [1, w], j ∈ [1, h], s are used for record window size, and r is used to record SVM caused by the window there are Aircraft TargetsResponse.When initialization, which is initialized as a full 0 matrix identical with image size.

Using the local pixel maximum of some candidate region as the center of the candidate region, referred to as: seed point (that is: may be usedThere can be the regional center position of Aircraft Targets).

(2) invalid seed point is rejected, most probable aircraft region is retained

For the accuracy rate for improving screening, the present invention proposes a kind of secondary windows method, specific as follows:

(A) first time image block window is determined

The window is a rectangular window centered on seed point, and the size of window is h₁×h₁, in which:

Wherein, int () function representation round.Original image is intercepted with the window, the figure that will be truncated toAs block is scaled a₃×a₃Size extracts the HOG feature of image block, then HOG feature is updated in SVM classifier, calculatesSVM classifier response r₁。

(B) second of image block window is determined

The window is a rectangular window centered on seed point, and the size of window is h₂×h₂, in which:

h₂=2 (h₁-1)+1 (10)

Original image is intercepted with the window, the image block being truncated to is scaled a₃×a₃Size extracts image blockThen HOG feature is updated in SVM classifier by HOG feature, calculate SVM classifier response r₂。

(C) judge whether to retain the region

If the classifier response being calculated twice is respectively less than 0, then it is assumed that the corresponding aircraft candidate regions of the seed pointDomain is invalid, therefore rejects the invalid seed point.

Otherwise it is assumed that the region that window includes there are aircraft or contains most of fuselage of aircraft, this can be rung twiceFor the larger value assignment answered into the response matrix on the corresponding coordinate position of seed point, which is known as the seed of response matrixPoint, and record the corresponding window size of the response.Specifically, if changing coordinates are as follows: (u, v), the svm classifier that needs are recordedDevice response is assigned to the r value at the seed point of response matrix in the two-dimensional array of element M (u, v)；It will generate corresponding when the responseInterception window size in square side length value, be assigned at the seed point of response matrix in the two-dimensional array of element M (u, v)S value.

(3) optimize response matrix, determine aircraft region

(A) optimize the window size recorded in each seed point

To the seed point of each response matrix, transformed edge of window long value h is first calculated according to the following formula₃:

Centered on the coordinate at the seed point of response matrix, with h₃×h₃Interception window is established to original image for window sizeAs being intercepted, the image block being truncated to is scaled a₃×a₃Size extracts the HOG feature of image block, then by HOG featureIt is updated in SVM classifier, calculates SVM classifier response r₃。

Transformed edge of window long value h is calculated further according to following formula₄:

Centered on the coordinate at the seed point of response matrix, with h₄×h₄Interception window is established to original image for window sizeAs being intercepted, the image block being truncated to is scaled a₃×a₃Size extracts the HOG feature of image block, then by HOG featureIt is updated in SVM classifier, calculates SVM classifier response r₄。

Find out r, r₃,r₄In maximum value, replace at the seed point of response matrix original r value in the two-dimensional array of element,With original s value in the two-dimensional array of element at the seed point of the corresponding edge of window long value replacement response matrix of the new r value.

(B) it determines aircraft region, obtains final target detection result

According to the response matrix after optimization, the coordinate recorded using at the seed point of response matrix is as aircraft mesh in original imageThe center of square area where mark, square region where the s value recorded using at the seed point as Aircraft Targets in original imageThe side length in domain can determine each aircraft region, obtain final Aircraft Targets testing result.

Claims (1)

Translated fromChinese

1.一种基于图像显著性与SVM的飞机目标检测方法，其特征在于步骤如下：1. an aircraft target detection method based on image salience and SVM, is characterized in that the steps are as follows:步骤1、数据准备：从公众可获得的各类图像中选取可见光机场图像作为训练样本及待检测目标图集，并转换为灰度图，图中飞机的最大外接正方形的尺寸从最小值：a₁×a₁，至最大值为a₂×a₂；选取部分图像作为训练样本，将图像中飞机区域切割后作为正样本，其他区域切割后作为负样本；剩余的部分图像作为测试样本；所述的训练的样本缩放为a₃×a₃大小的尺寸；Step 1. Data preparation: Select visible light airport images from various types of images available to the public as training samples and target atlases to be detected, and convert them into grayscale images. The size of the largest circumscribed square of the aircraft in the figure is from the smallest value: a₁ × a₁ , and the maximum value is a₂ ×a₂ ; select some images as training samples, cut the aircraft area in the image as positive samples, and cut other areas as negative samples; the remaining part of the images are used as test samples; all The training samples described above are scaled to a₃ × a₃ size;步骤2：分别提取正负训练样本的HOG特征；Step 2: Extract HOG features of positive and negative training samples respectively;步骤3：将提取到的每一个正负样本的HOG特征与其类别标签，正样本为1，负样本为-1组合成一个向量，训练SVM分类器；Step 3: Combine the extracted HOG feature of each positive and negative sample and its category label, the positive sample is 1, and the negative sample is -1 into a vector, and train the SVM classifier;步骤4：采用基于剩余谱理论的显著性检测方法获取测试样本中图像的显著图；Step 4: use the saliency detection method based on residual spectrum theory to obtain the saliency map of the images in the test sample;步骤5：通过面积阈值化方法提取显著图中的连通域，产生飞机目标的候选区域，过程如下：Step 5: Extract the connected domain in the saliency map by the area thresholding method, and generate the candidate area of the aircraft target. The process is as follows:步骤5a：首先计算显著图的均值和方差Step 5a: First compute the mean of the saliency map and variance其中，w表示图像的宽，h表示图像的高，Sal(i,j)表示显著图中第i行第j列的像素值；Among them, w represents the width of the image, h represents the height of the image, and Sal(i,j) represents the pixel value of the i-th row and the j-th column in the saliency map;步骤5b：利用显著图的均值和方差计算阈值T，以该阈值T分割显著图；所述阈值T：Step 5b: Use the mean of the saliency map and variance Calculate the threshold T, and segment the saliency map with the threshold T; the threshold T:系数k是为了平衡标准差和均值在阈值T取值时的权重The coefficient k is to balance the standard deviation and mean The weight when the threshold value T is taken步骤5c、对阈值化后的显著图进行连通域滤波：以大小为n₁×n₁的矩形窗，以第i行第j列像素(i,j)为中心截取一个邻域；以某区域的像素个数来定义该区域的面积，求出处于所述邻域范围内的显著区域的面积；如果该面积大于一定阈值，就保留该区域，否则不保留；将所有保留下来的区域作为飞机目标的候选区域；Step 5c. Perform connected domain filtering on the thresholded saliency map: take a rectangular window of size n₁ ×n₁ , and intercept a neighborhood with the pixel (i, j) in the i-th row and the j-th column as the center; The area of the area is defined by the number of pixels in the area, and the area of the significant area within the neighborhood range is obtained; if the area is greater than a certain threshold, the area is reserved, otherwise it is not reserved; all reserved areas are used as aircraft The candidate region of the target;步骤6：从飞机目标的候选区域提取飞机目标，过程如下：Step 6: Extract the aircraft target from the candidate area of the aircraft target, the process is as follows:步骤6a、构建响应矩阵：Step 6a, construct the response matrix:构建一个二维的响应矩阵M，矩阵中每个元素的取值为一个二维数组M_i,j(s,r)其中i∈[1,w],j∈[1,h]，s用于记录窗口尺寸,r用于记录存在飞机目标的窗口所产生的SVM响应；初始化时，该响应矩阵初始化为一个与图像大小相同的全0矩阵；Construct a two-dimensional response matrix M, and the value of each element in the matrix is a two-dimensional array M_i,j (s,r) where i∈[1,w],j∈[1,h], s uses For recording the window size, r is used to record the SVM response generated by the window where the aircraft target exists; during initialization, the response matrix is initialized to an all-zero matrix with the same size as the image;以任一候选区域的局部像素极大值作为该候选区域的中心，称为：种子点；Taking the local pixel maximum value of any candidate region as the center of the candidate region, it is called: the seed point;步骤6b、二次窗口法剔除无效种子点，保留最可能的飞机区域，过程如下：Step 6b, the quadratic window method removes invalid seed points and retains the most likely aircraft area, the process is as follows:(A)确定第一次图像块窗口：窗口是一个以种子点为中心的矩形窗口，窗口的尺寸为h₁×h₁，其中：(A) Determine the first image block window: the window is a rectangular window centered on the seed point, and the size of the window is h₁ ×h₁ , where:其中，int()函数表示四舍五入取整；以该窗口对原图像进行截取，将截取到的图像块缩放为a₃×a₃大小，提取图像块的HOG特征，然后将HOG特征代入到SVM分类器中，计算SVM分类器响应值r₁；Among them, the int() function indicates rounding; use this window to intercept the original image, scale the intercepted image block to a₃ × a₃ size, extract the HOG feature of the image block, and then substitute the HOG feature into the SVM classification In the classifier, calculate the SVM classifier response value r₁ ;(B)确定第二次图像块窗口：窗口是一个以种子点为中心的矩形窗口，窗口的尺寸为h₂×h₂，其中：(B) Determine the second image block window: the window is a rectangular window centered on the seed point, and the size of the window is h₂ ×h₂ , where:h₂＝2(h₁-1)+1h₂ =2(h₁ -1)+1以该窗口对原图像进行截取，将截取到的图像块缩放为a₃×a₃大小，提取图像块的HOG特征，然后将HOG特征代入到SVM分类器中，计算SVM分类器响应值r₂；Use this window to intercept the original image, scale the intercepted image block to a₃ × a₃ size, extract the HOG feature of the image block, and then substitute the HOG feature into the SVM classifier, and calculate the SVM classifier response value r₂ ;(C)判断是否保留该区域：(C) Determine whether to reserve the area:如果两次计算得到的分类器响应值均小于0，则认为该种子点对应的飞机候选区域无效，因此剔除该无效种子点；If the classifier response value obtained by the two calculations is less than 0, the aircraft candidate area corresponding to the seed point is considered invalid, so the invalid seed point is eliminated;否则，认为窗口包含的区域存在飞机或包含了飞机的大部分机身，将该两次响应中的较大值赋值到响应矩阵中种子点对应的坐标位置上，称为响应矩阵的种子点，并记录下该响应对应的窗口尺寸；具体地，设当前坐标为：(u,v)，将需要记录的SVM分类器响应值，赋给响应矩阵的种子点处元素M(u,v)的二维数组中的r值；将产生该响应时对应的截取窗口尺寸中的正方形边长值，赋给响应矩阵的种子点处元素M(u,v)的二维数组中的s值；Otherwise, it is considered that the area contained in the window contains an aircraft or most of the fuselage of the aircraft, and the larger value of the two responses is assigned to the coordinate position corresponding to the seed point in the response matrix, which is called the seed point of the response matrix. And record the window size corresponding to the response; specifically, set the current coordinates as: (u, v), assign the SVM classifier response value to be recorded to the element M(u, v) at the seed point of the response matrix The r value in the two-dimensional array; the square side length value in the corresponding interception window size when the response is generated is assigned to the s value in the two-dimensional array of the element M(u, v) at the seed point of the response matrix;步骤6c、优化响应矩阵，确定飞机区域：Step 6c, optimize the response matrix to determine the aircraft area:(1)优化每个种子点中记录的窗口尺寸：(1) Optimize the window size recorded in each seed point:对每一个响应矩阵的种子点，先根据以下公式计算变换后的窗口边长值h₃：For each seed point of the response matrix, first calculate the transformed window side length value h₃ according to the following formula:以响应矩阵的种子点处的坐标为中心，以h₃×h₃为窗口尺寸建立截取窗口对原图像进行截取，将截取到的图像块缩放为a₃×a₃大小，提取图像块的HOG特征，然后将HOG特征代入到SVM分类器中，计算SVM分类器响应值r₃；Taking the coordinates of the seed point of the response matrix as the center, establishing an interception window with h₃ ×h₃ as the window size to intercept the original image, scaling the intercepted image block to a₃ ×a₃ size, and extracting the HOG of the image block feature, and then substitute the HOG feature into the SVM classifier, and calculate the SVM classifier response value r₃ ;再根据以下公式计算变换后的窗口边长值h₄：Then calculate the transformed window side length value h₄ according to the following formula:以响应矩阵的种子点处的坐标为中心，以h₄×h₄为窗口尺寸建立截取窗口对原图像进行截取，将截取到的图像块缩放为a₃×a₃大小，提取图像块的HOG特征，然后将HOG特征代入到SVM分类器中，计算SVM分类器响应值r₄；Taking the coordinates of the seed point of the response matrix as the center, establishing an interception window with h₄ ×h₄ as the window size to intercept the original image, scaling the intercepted image block to a₃ ×a₃ size, and extracting the HOG of the image block feature, and then substitute the HOG feature into the SVM classifier, and calculate the SVM classifier response value r₄ ;找出r,r₃,r₄中的最大值，替换响应矩阵的种子点处元素的二维数组中原有的r值，以新的r值对应的窗口边长值替换响应矩阵的种子点处元素的二维数组中原有的s值；Find the maximum value of r, r₃ , r₄ , replace the original r value in the two-dimensional array of elements at the seed point of the response matrix, and replace the seed point of the response matrix with the window side length value corresponding to the new r value The original s value in the two-dimensional array of elements;确定飞机区域，得到最终目标检测结果；Determine the aircraft area and get the final target detection result;(2)依据优化后的响应矩阵，以响应矩阵的种子点处记录的坐标作为原图像中飞机目标所在正方形区域的中心，以该种子点处记录的s值作为原图像中飞机目标所在正方形区域的边长，即可确定每个飞机区域，得到最终的飞机目标检测结果。(2) According to the optimized response matrix, the coordinates recorded at the seed point of the response matrix are used as the center of the square area where the aircraft target is located in the original image, and the s value recorded at the seed point is used as the square area where the aircraft target is located in the original image. The side length of each aircraft area can be determined, and the final aircraft target detection result can be obtained.