CN108090485A - Display foreground extraction method based on various visual angles fusion - Google Patents

Abstract

The invention discloses an image foreground automatic extraction method based on multi-view fusion, which mainly solves the problems of cumbersome extraction process and inaccurate extraction of foreground edges in the prior art. The implementation plan is: first train the SVM classifier, and then obtain the grayscale image of the image to be extracted; detect the sub-image containing the foreground in the gray-scale image through the trained SVM classifier; put the sub-image in the image to be extracted As the input of the GrabCut algorithm, foreground extraction is performed on the image to be extracted, and the extraction result under the pixel perspective of the image to be extracted is obtained; the image under the superpixel perspective is generated from the image to be extracted by the SLIC algorithm; the image under the superpixel perspective and The extraction results under the pixel perspective are fused to obtain the foreground extraction result of the image to be extracted. The invention simplifies the foreground extraction process, improves extraction efficiency and precision, and can be used for stereo vision, image semantic recognition, three-dimensional reconstruction and image search.

Description

Automatic extraction method of image foreground based on multi-view fusion

technical field

The invention belongs to the technical field of image processing, and further relates to an image foreground automatic extraction method based on multi-view fusion. The invention can be used in the application and research of stereo vision, image semantic recognition, image search, etc.

Background technique

Foreground extraction is a means to extract objects of interest in an image. It divides the image into several specific regions with unique properties and proposes the technology and process of the target of interest, and has become a key step from image processing to image analysis. The specific explanation is to divide the image into several complementary and overlapping regions according to features such as gray scale, color, texture and shape, and make these features appear similar in the same region, but show obvious differences in different regions. After decades of development and changes, foreground extraction has gradually formed its own scientific system, and new extraction methods have emerged one after another. It has become an interdisciplinary field and has attracted extensive attention from researchers and application personnel in various fields. Such as the field of medicine, aerospace remote sensing, industrial testing, security and military fields, etc.

Current foreground extraction methods mainly include threshold-based foreground extraction methods, edge-based foreground extraction methods, region-based foreground extraction methods, graph cutting-based foreground extraction methods, energy functional-based foreground extraction methods, and image foreground extraction methods based on deep learning. extraction method, etc. Among them, the foreground extraction method based on graph cutting is favored because of its high extraction accuracy and simple operation. The foreground extraction method based on graph cutting is a combined optimization method based on graph theory. According to the user's interaction information, it maps an image into A network graph, and establish an energy function about the label, use the maximum flow minimum cut algorithm to cut the network graph iteratively for a limited number of times, and obtain the minimum cut of the network graph as the foreground extraction result of the image. However, due to the existence of human-computer interaction, when extracting multiple images, the amount of manual operation is too large, which limits its application in engineering. For example, "GrabCut in One Cut" published by Meng Tang et al. at the 2013 IEEE International Conference on Computer Vision in 2013, the user selects the foreground area, and then maps the area where the foreground is located into a graph, and performs a limited number of iterative cutting on the map through One Cut , to obtain the foreground extraction result of the image, but human-computer interaction is required to calibrate the foreground area, resulting in a cumbersome foreground extraction process, and limited energy iterative optimization can only obtain the minimum cut of a better solution, and it is difficult to obtain accurate foreground edges.

Contents of the invention

The purpose of the present invention is to address the above-mentioned deficiencies in the prior art, and propose an image foreground automatic extraction method based on multi-view fusion, which is used to solve the problems caused by the existence of human-computer interaction in the existing foreground extraction method based on graph cutting. The foreground extraction process is relatively cumbersome and the problem of inaccurate foreground edges caused by limited energy iterative optimization.

In order to achieve the above object, the technical scheme that the present invention takes includes as follows:

(1) train the SVM classifier to obtain the trained SVM classifier;

(2) Grayscale the image to be extracted to obtain a grayscale image;

(3) through the trained SVM classifier, detect the sub-image p _k that contains the foreground target in the grayscale image;

(3a) Use multi-scale windows to slide line by line in the grayscale image according to the set interval, and obtain an image set P={p ₁ ,p ₂ ,...p _k ,... ,p _q }, where, k∈[1,q], p _k is the kth sub-image, and q is the number of sub-images;

(3b) extracting the histogram HOG feature of the direction gradient of each sub-image p _k in the image set P, and inputting it into the trained SVM classifier for classification, and calculating the label l _pk of the sub-image p _k ;

(3c) Determine whether the label l _pk of the sub-image p _k is positive, if so, the sub-image p _k contains the foreground object, record the position of the sub-image p _k in the image to be extracted, that is, the pixel in the upper left corner of the sub-image p _k is to be extracted The corresponding position (x _min , y _min ) of the extracted image and the pixel in the lower right corner are at the corresponding position (x _max , y _max ) of the image to be extracted, perform step (4), otherwise, discard the image p _k ;

(4) Foreground extraction of the image to be extracted:

Using the pixel in the upper left corner of the sub-image p _k at the corresponding position (x _min , y _min ) of the image to be extracted and the pixel in the lower right corner at the corresponding position (x _max , y _max ) of the image to be extracted, the human-computer interaction of the GrabCut algorithm Perform replacement, and use the replacement result to extract the foreground of the image to be extracted, and obtain the extraction result S ₁ (x, y) under the pixel perspective of the image to be extracted;

(5) Use the simple linear iterative clustering algorithm SLIC to calculate the superpixels of the image to be extracted, and obtain the image under the superpixel perspective: B={b ₁ ,b ₂ ,..., _bi ,...,b _m } , i∈[1,m], b _i is the i-th superpixel, m is the number of superpixels;

(6) Perform multi-view fusion on the image B under the superpixel perspective and the extraction result S ₁ (x, y) under the pixel perspective of the image to be extracted to obtain the foreground S ₂ ( _xi , y _i ) of the image to be extracted.

Compared with the prior art, the present invention has the following advantages:

1) The present invention adopts the trained SVM classifier to obtain the sub-image where the foreground is located in the image to be extracted, and uses the position coordinates of the sub-image in the image to be extracted to replace the rectangular area obtained by the human-computer interaction of the GrabCut algorithm as an algorithm input to realize the treatment The foreground extraction of the extracted image fully combines the SVM classifier and the GrabCut algorithm, which can automatically complete the image foreground extraction process, and solves the cumbersome foreground extraction process caused by the existence of human-computer interaction in the existing foreground extraction method based on graph cutting problem, effectively improving the efficiency of image foreground extraction.

2) The present invention uses the SLIC algorithm to extract superpixels from the image to be extracted, fully utilizes the characteristics of good consistency in the superpixel block, and fuses the image under the superpixel perspective and the extraction result under the pixel perspective to obtain Extract the foreground of the image with accurate extraction results,

3) The present invention makes the foreground extraction result more accurate and smooth by introducing superpixels, solves the problem of inaccurate foreground edges caused by limited energy iterative optimization in the existing foreground extraction method based on graph cutting, and improves the image quality. Accuracy of foreground extraction.

Description of drawings

Fig. 1 is the realization flowchart of the present invention;

Fig. 2 is a sample image set structure diagram in the present invention;

Fig. 3 is a schematic diagram of extracting HOG features in the present invention;

Fig. 4 is a visual display diagram of HOG features in the present invention;

Fig. 5 is a diagram of the extraction experiment results of pedestrians and leaves as the foreground in the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

With reference to Fig. 1, the image foreground automatic extraction method based on multi-view fusion comprises the following steps:

Step 1, train the SVM classifier.

(1a) Collect a sample image set containing foreground categories, and grayscale all the sample images therein to obtain a sample grayscale image set;

The structure diagram of the sample image set containing the foreground category is shown in Figure 2, which includes positive samples, negative samples and sample label files, where positive samples are images that contain foreground, negative samples are images that do not contain foreground, and sample label files use It is used to explain the categories and storage locations of positive samples and negative samples;

The grayscale of all the sample images in the sample image set is to carry out weighted average of the red component R, green component G, and blue component B of the three channels in the sample image to obtain the grayscale image of the sample image. Degree value Gray:

Gray＝R×0.299+G×0.587+B×0.114;

(1b) Extract the HOG feature of the direction gradient histogram of each image in the sample grayscale image set:

Referring to Figure 3, the specific implementation of this step is as follows:

(1b1) Divide the input image into several connected and non-overlapping units, and calculate the gradient magnitude G(x, y) and gradient direction α(x, y) of the pixel in each unit, and the calculation formulas are respectively :

Among them, G _x (x, y) = H (x + 1, y) - H (x-1, y) represents the horizontal gradient at the pixel point (x, y) in the input image, G _y (x, y )=H(x,y+1)-H(x,y-1) indicates the vertical direction gradient at the pixel point (x,y) in the input image, and H(x,y) indicates the pixel point (x,y) in the input image , the pixel value at y);

(1b2) Divide all gradient directions α(x, y) into 9 angles as the horizontal axis of the histogram, and accumulate the gradient values corresponding to each angle range as the vertical axis of the histogram to obtain a gradient histogram;

(1b3) Count the gradient histogram of each unit to obtain the feature descriptor of each unit;

(1b4) 8×8 units are formed into a block, and the feature descriptors of all units in a block are connected in series to obtain the HOG feature descriptor of the block's direction gradient histogram;

(1b5) Concatenate the HOG feature descriptors of all blocks in the input image to obtain the HOG feature of the directional gradient histogram of the input image, and the visual display of the HOG feature of the directional gradient histogram is shown in Figure 4. Among them, Figure 4(a) is an example image, and Figure 4(b) is the HOG feature map of the oriented gradient histogram of the example image. It can be seen from Figure 4 that the HOG feature of the oriented gradient histogram can well describe the local area through gradient or edge direction density. the appearance and shape of the target;

(1b6) Concatenate the histogram HOG features of all input images in the sample grayscale image set to obtain the HOG feature set of the histogram of orientation gradients in the sample grayscale image set;

(1c) Use all the HOG features of the histogram of orientation gradients in the HOG feature set to train the SVM classifier to obtain a trained SVM classifier.

Step 2: grayscale the image to be extracted to obtain a grayscale image.

The red component R', the green component G', and the blue component B' of the three channels in the image to be extracted are weighted and averaged to obtain the gray value Gray' of each pixel in the image to be extracted:

Gray'=R'×0.299+G'×0.587+B'×0.114

The grayscale image of the image to be extracted is obtained according to the grayscale value of each pixel.

Step 3, use the trained SVM classifier to detect the sub-image p _i containing the foreground object in the gray image of the image to be extracted, and obtain the pixel in the upper left corner of the sub-image p _i at the corresponding position of the image to be extracted (x _min , y _min ) and the pixel in the lower right corner is at the corresponding position (x _max , y _max ) of the image to be extracted.

(3a) Use multi-scale windows to slide line by line in the grayscale image of the image to be extracted according to the set interval, and obtain an image set composed of multiple sub-images: P={p ₁ ,p ₂ ,...p _k ,...,p _q }, where p _k is the kth sub-image, and q is the number of sub-images;

(3b) Extract the HOG feature of the histogram of oriented gradients of each sub-image p _k in the image set P, and input it into the trained SVM classifier for classification, and obtain the label l _pk of the sub-image p _k , and the calculation of l _pk The formula is:

in, _k is the kth normal vector of the hyperplane of the SVM classifier, x _k is the HOG feature of the directional gradient histogram of the sub-image 0 _k , k∈[1,q], q is the number of sub-images, φ is the SVM classifier The displacement term of the hyperplane.

(3c) Determine whether the label l _pk of the sub-image p _k is positive, if so, the sub-image p _k contains the foreground object, record the position of the sub-image p _k in the image to be extracted, that is, the pixel in the upper left corner of the sub-image p _k is to be extracted The corresponding position (x _min , y _min ) of the extracted image and the pixel in the lower right corner are in the corresponding position (x _max , y _max ) of the image to be extracted, and the pixel in the upper left corner of the sub-image p _i is in the corresponding position (x max , y max ) of the image to be extracted _min , y _min ) and the corresponding position (x _max , y _max ) of the lower right corner of the image to be extracted to form a rectangular area containing the foreground, go to step 4, otherwise, discard the sub-image p _k .

Step 4, perform foreground extraction on the image to be extracted.

Foreground extraction methods mainly include threshold-based foreground extraction methods, edge-based foreground extraction methods, region-based foreground extraction methods, graph cutting-based foreground extraction methods, energy functional-based foreground extraction methods, and image foreground extraction based on deep learning. method etc. This example uses but is not limited to the foreground extraction method based on graph cutting to perform foreground extraction based on the GrabCut algorithm. The specific implementation is: use the pixels in the upper left corner and lower right corner of the sub-image p _i to be at the corresponding position of the image to be extracted (x _min , y _min ), (x _max , y _max ) and the image to be extracted are used as the input of the GrabCut algorithm, the foreground is extracted from the image to be extracted, and the extraction result S ₁ (x, y) under the pixel perspective of the image to be extracted is obtained.

Step 5, calculate the superpixels of the image to be extracted.

To calculate the superpixels of the image to be extracted, the existing methods include methods based on graph theory and methods based on gradient descent; this example uses but is not limited to methods based on gradient descent to extract superpixels based on the SLIC algorithm; the specific steps are as follows:

(5a) convert the image to be extracted to the CIE-Lab color space from the RGB color space, and obtain the CIE-Lab image;

The image to be extracted is converted from RGB color space to CIE-Lab color space, wherein there is no direct conversion formula between RGB and LAB, it must use XYZ color space as the intermediate layer, and the conversion formula is:

Among them, r, g, b are the three channel components of the image pixel to be extracted, R, G, B are the three channel components of r, g, b after correction by the correction function gamma(t),

The conversion formula from the RGB color space to the XYZ color space obtains the X, Y, Z three channel components of the XYZ color space; the conversion formula is:

in, is a 3×3 matrix;

In the CIE-Lab color space, the conversion formula from the XYZ color space to the CIE-Lab color space obtains the values of the three channels of the CIE-Lab color space L, a, and b, and the conversion formula is:

L=116f(Y/Y _n )

b=500(f(X/X _n )-f(Y/Y _n ))

a=200(f(Y/Y _n )-f(Z/Z _n ))

Among them, X, Y, and Z are the three-channel components converted from the RGB color space to the XYZ color space; the values of X _n , Y _n , and Z _n are respectively 95.047, 100.0, and 108.883; f(X/X _n ), f( Y/Y _n ), f(Z/Z _n ) is calculated by the following function:

(5b) Initialize the clustering centers of superpixels: set the number of superpixels m=200, distribute the superpixel clustering centers evenly according to the number of superpixels in the CIE-Lab image, and obtain the clustering center set in, is the i-th cluster center after the d-th iteration, a total of m, where, l _i , a _i , b _i are the three channels of the CIE-Lab color space, (xi _, y _i ) are the coordinates of b _i ;

(5c) For each pixel pixel of the CIE-Lab image, set label l(pixel)=-1 and distance d(pixel)=∞;

(5d) Calculate the cluster center in the cluster center set C ^d respectively Gradient values of all pixels in the 3×3 neighborhood of , and move the cluster center to the pixel with the smallest gradient in this area to obtain a new cluster center set C ^d+1 ;

(5e) For each cluster center in the cluster center set C ^d For each pixel pixel in 2S×2S=[l _p , a _p , b _p , x _p , y _p ], calculate Distance D(pixel) from pixel:

in is the color difference between pixels,

is the spatial distance between pixels,

M is the maximum value of d _c , N is the number of all pixels in the image, and m is the number of superpixels set;

(5f) Compare the size of d(pixel) and D(pixel), if D(pixel)<d(pixel), then assign D(pixel) to d(pixel), set d(pixel)=D(pixel) , that is, use d(pixel) to record the pixel to the cluster center , and use the pixel label l(pixel) to mark the pixel as belonging to the ith superpixel, l(pixel)=i, to obtain a new superpixel b _i ;

(5g) Repeat steps (5d)~(5f) to update the cluster centers until the residual error converges, and obtain the superpixel image B={b ₁ ,b ₂ ,..., _bi ,...,b _m }.

Step 6: Perform multi-view fusion on the image under the superpixel perspective and the extraction result S ₁ (x, y) under the pixel perspective of the image to be extracted to obtain the foreground S ₂ ( _xi , y _i ) of the image to be extracted.

(6a) Weight the label l _ij in the extraction result S ₁ (x, y) of all pixels contained in the superpixel b _i of the image B under the superpixel perspective, and obtain the label confidence of the superpixel b _i Degree Score _bi :

Score _bi = ∑ l _ij ;

(6b) Set the confidence threshold gate, compare the confidence threshold gate with the label confidence Score _bi of the superpixel b _i , obtain the label l _bi under the view angle of the superpixel b _i , and use the label l _bi as the pixel point The label S ₂ (xi _, y _i ) of ( _xi , y _i ), the labels of all pixels in the superpixel b _i are the same as the label of the superpixel, and S ₂ ( _xi , y _i ) is the image to be extracted Foreground, where ( _xi ,y _{) ∈bi} ;

The smaller the set confidence gate is, the smaller the probability that the superpixel _bi is judged as the foreground, and the larger the gate is, the greater the probability that the superpixel _{bi is} judged as the foreground, but when the gate is too large, the foreground extraction result There will be too much noise in the

The confidence threshold gate is compared with the label confidence Score _bi of the superpixel b _i to obtain the label l _bi of the superpixel b _i , and the comparison formula is:

Among them, l _bi is the label of superpixel b _i , num _bi is the number of pixels in superpixel b _i , gate is the confidence threshold, 1 is the foreground label, and 0 is the background label;

The multi-view fusion refers to fusing the image B under the superpixel perspective and the extraction result S ₁ (x, y) under the pixel perspective.

The technical effect of the present invention will be further described below by foreground extraction experiment:

1. Experimental conditions and content

The experiment of the present invention extracts the target of pedestrians and leaves respectively, the training data is the image set of pedestrians and leaves randomly found by the network, the number of images is respectively 736 and 186, and the positive and negative samples are respectively taken for each picture, and then labels are made, A sample image set containing the category of pedestrians and a sample image set containing the category of leaves are formed respectively.

By programming in MATLAB R2017a, the foreground extraction of pedestrians and leaves is realized, and the results are shown in Figure 5.

2. Analysis of experimental results:

It can be seen from Figure 5 that for the four images tested on the two types of data, the output foreground extraction results have no noise, and the extracted foreground edges are better. For example, for the foreground extraction results of the four images of the leaf category, the extracted foreground Edges are extremely accurate. It has better compatibility with the completeness of the foreground contained in the input image. For example, for the input image 3 of the pedestrian category, a bust image is used as the input image, and a good foreground extraction effect can still be obtained.

In addition, it can be seen from Fig. 5 that after the SVM classifier is trained, the present invention can automatically complete the automatic process of the foreground of the image to be extracted, obtain the foreground extraction result of the image to be extracted, and solve the problems in the existing foreground extraction method based on graph cutting. The problem that human-computer interaction is needed to assist the extraction, and the present invention makes full use of the characteristics of good consistency in the super-pixel block, and repairs the edge of the extraction result under the pixel perspective output by the GrabCut algorithm, so that the foreground extraction result is more accurate and smooth , to obtain accurate foreground extraction results and improve the accuracy of foreground extraction.

Claims (8)

1. An automatic image foreground extraction method based on multi-view fusion is characterized in that:

(1) training the SVM classifier to obtain a trained SVM classifier;

(2) graying an image to be extracted to obtain a grayscale image;

(3) detecting a sub-image p containing a foreground target in the gray image through a trained SVM classifier_k；

(3a) Adopting a multi-scale window to slide line by line in the gray level image according to a set interval to obtain a channelImage set P ═ { P } composed of a plurality of sub-images₁,p₂,...p_k,...,p_qIn which k ∈ [1, q ]],p_kQ is the number of sub-images;

(3b) extracting each sub-image P in image set P_kThe HOG features of the histogram of directional gradients are input into a trained SVM classifier for classification, and a sub-image p is obtained through calculation_kLabel l of_pk；

(3c) Judging the subimage p_kLabel l of_pkIf positive, sub-picture p_kContaining foreground objects, recording subimages p_kAt the position of the image to be extracted, i.e. subimage p_kThe pixel at the upper left corner is at the corresponding position (x) of the image to be extracted_min,y_min) And the pixel at the lower right corner at the corresponding position (x) of the image to be extracted_max,y_max) Executing the step (4), otherwise, discarding the image p_k；

(4) Carrying out foreground extraction on an image to be extracted:

using subimages p_kThe pixel at the upper left corner is at the corresponding position (x) of the image to be extracted_min,y_min) And the pixel at the lower right corner at the corresponding position (x) of the image to be extracted_max,y_max) Replacing the human-computer interaction of the GrabConut algorithm, and performing foreground extraction on the image to be extracted by using the replacement result to obtain an extraction result S under the pixel viewing angle of the image to be extracted₁(x,y)；

(5) Calculating the super pixels of the image to be extracted by adopting a simple linear iterative clustering algorithm SLIC to obtain the image under the super pixel viewing angle: b ═ B₁,b₂,...,b_i,...,b_m}，i∈[1,m],b_iIs the ith super pixel, and m is the number of super pixels;

(6) extracting result S of image B under super pixel visual angle and image to be extracted under pixel visual angle₁(x, y) carrying out multi-view fusion to obtain the foreground S of the image to be extracted₂(x_i,y_i)。

2. The method of claim 1, wherein the training of the SVM classifier in step (1) is performed by:

(1a) collecting a sample image set containing a foreground category, and graying all sample images in the sample image set to obtain a sample gray image set;

(1b) extracting HOG (histogram of oriented gradient) features of each image in the sample gray level image set to obtain a HOG feature set of the sample HOG;

(1c) and training the SVM classifier by adopting all HOG characteristics in the HOG characteristic set of the sample HOG to obtain the trained SVM classifier.

3. The method of claim 2, wherein: graying all sample images in the sample image set in the step (1a), namely performing weighted average on three channels of red components R, green components G and blue components B in the sample images to obtain a Gray value Gray of the sample Gray image:

Gray＝R×0.299+G×0.587+B×0.114。

4. the method of claim 2, wherein: in the step (1b), the HOG characteristic of each image in the sample gray level image set is extracted, and the method comprises the following steps:

(1b1) dividing an input image into a plurality of units which are connected and adjacent without overlapping, and calculating the gradient amplitude G (x, y) and the gradient direction α (x, y) of a pixel in each unit:

<mrow> <mi>G</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>)</mo> </mrow> <mo>=</mo> <msqrt> <mrow> <msub> <mi>G</mi> <mi>x</mi> </msub> <msup> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>+</mo> <msub> <mi>G</mi> <mi>y</mi> </msub> <msup> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>)</mo> </mrow> <mn>2</mn> </msup> </mrow> </msqrt> </mrow>

<mrow> <mi>&amp;alpha;</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>)</mo> </mrow> <mo>=</mo> <msup> <mi>tan</mi> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msup> <mrow> <mo>(</mo> <mfrac> <mrow> <msub> <mi>G</mi> <mi>x</mi> </msub> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>)</mo> </mrow> </mrow> <mrow> <msub> <mi>G</mi> <mi>y</mi> </msub> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>)</mo> </mrow> </mrow>

wherein G is_x(x,y)＝H(x+1,y)-H(x-1,y)，G_y(x, y) ═ H (x, y +1) -H (x, y-1) respectively represent the horizontal direction gradient and the vertical direction gradient at the pixel point (x, y) in the input image, H (x +1, y) represents the pixel value at the pixel point (x +1, y) in the input image, H (x-1, y) represents the pixel value at the pixel point (x-1, y) in the input image, H (x, y +1) represents the pixel value at the pixel point (x, y +1) in the input image, and H (x, y-1) represents the pixel value at the pixel point (x, y-1) in the input image;

(1b2) dividing all gradient directions α (x, y) into 9 angles as the horizontal axis of the histogram, and accumulating gradient values corresponding to each angle range as the vertical axis of the histogram to obtain a gradient histogram;

(1b3) counting the gradient histogram of each unit to obtain a feature descriptor of each unit;

(1b4) combining n × n units into a block, and connecting the feature descriptors of all the units in the block in series to obtain the HOG feature descriptor of the block;

(1b5) connecting direction gradient histogram HOG feature descriptors of all blocks in the input image in series to obtain the HOG feature of the direction gradient histogram of the input image;

(1b6) and connecting the HOG features of the direction gradient histograms of all the input images in the sample gray level image set in series to obtain a HOG feature set of the direction gradient histograms of the sample gray level image set.

5. The method of claim 1, wherein: in step (3b) the operator image p is calculated_kLabel l of_pkBy the following formula:

wherein,the k-th normal vector, x, of the hyperplane of the SVM classifier_kAs a subimage p_kHOG feature of histogram of oriented gradients, k ∈ [1, q ]]Q is the number of sub-images and phi is the displacement term of the hyperplane of the SVM classifier.

6. The method of claim 1, wherein: calculating the superpixel of the image to be extracted in the step (5), wherein the method comprises the following implementation steps:

(5a) converting an image to be extracted from an RGB color space to a CIE-Lab color space to obtain a CIE-Lab image;

(5b) initializing the cluster centers of the superpixels: setting the number of superpixels, and uniformly distributing superpixel clustering centers according to the number of the superpixels in the CIE-Lab image to obtain a clustering center setWherein,m cluster centers after the d iteration, wherein l_i,a_i,b_iThree channels of CIE-Lab color space, (x)_i,y_i) Is b is_iCoordinates;

(5c) setting a label l (pixel) ═ 1 and a distance d (pixel) ∞foreach pixel of the CIE-Lab image;

(5d) separately computing a cluster center set C^dCenter of clusterAnd moving the clustering center to the pixel point with the minimum gradient in the field to obtain a new clustering center set C^d+1；

(5e) For cluster center set C^dEach cluster center inEach pixel in 2 sx 2S ═ l_p,a_p,b_p,x_p,y_p]CalculatingDistance D (pixel) from pixel:

<mrow> <mi>D</mi> <mrow> <mo>(</mo> <mi>p</mi> <mi>i</mi> <mi>x</mi> <mi>e</mi> <mi>l</mi> <mo>)</mo> </mrow> <mo>=</mo> <msqrt> <mrow> <msubsup> <mi>d</mi> <mi>c</mi> <mn>2</mn> </msubsup> <mo>+</mo> <msup> <mrow> <mo>(</mo> <mfrac> <msub> <mi>d</mi> <mi>s</mi> </msub> <mi>S</mi> </mfrac> <mo>)</mo> </mrow> <mn>2</mn> </msup> <msup> <mi>M</mi> <mn>2</mn> </msup> </mrow> </msqrt> </mrow>

whereinThe difference in color between the pixels is determined,

the spatial distance between the pixels is the distance between the pixels,

m is d_cThe maximum value of (a) is,n is the number of all pixel points of the image, and m is the set number of super pixels;

(5f) comparing the size of D (pixel) with that of D (pixel), if D (pixel) < D (pixel), D (pixel) is assigned to D (pixel), and l (pixel) < i, a new super pixel b is obtained_i；

(5g) Continuously executing the steps (5d) - (5f), updating the clustering center until the residual error is converged, and obtaining the superpixel image B ═ B₁,b₂,...,b_i,...,b_m}。

7. The method according to claim 1, wherein the step (6) comprises extracting the image B under the super-pixel view angle and the extraction result S under the pixel view angle of the image to be extracted₁(x, y) performing multi-view fusion as follows:

(6a) for the super pixel B of the image B under the super pixel view angle_iExtracting result S of all contained pixels under pixel viewing angle₁Label l in (x, y)_ijWeighting to obtain super pixel b_iTag confidence Score of_bi；

(6b) Setting a confidence threshold gate, and connecting the confidence threshold gate with the super-pixel b_iTag confidence Score of_biComparing to obtain the super pixel b_iLabel under visual angle_bi：

<mrow> <msub> <mi>l</mi> <mrow> <mi>b</mi> <mi>i</mi> </mrow> </msub> <mo>=</mo> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mn>1</mn> </mtd> <mtd> <mrow> <mi>i</mi> <mi>f</mi> </mrow> </mtd> <mtd> <mrow> <msub> <mi>Score</mi> <mrow> <mi>b</mi> <mi>i</mi> </mrow> </msub> <mo>&gt;</mo> <msub> <mi>num</mi> <mrow> <mi>b</mi> <mi>i</mi> </mrow> </msub> <mo>/</mo> <mi>g</mi> <mi>a</mi> <mi>t</mi> <mi>e</mi> </mrow> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mrow> <mi>i</mi> <mi>f</mi> </mrow> </mtd> <mtd> <mrow> <msub> <mi>Score</mi> <mrow> <mi>b</mi> <mi>i</mi> </mrow> </msub> <mo>&lt;</mo> <msub> <mi>num</mi> <mrow> <mi>b</mi> <mi>i</mi> </mrow> </msub> <mo>/</mo> <mi>g</mi> <mi>a</mi> <mi>t</mi> <mi>e</mi> </mrow> </mtd> </mtr> </mtable> </mfenced> </mrow>

Wherein, num_biIs a super pixel b_iThe number of middle pixels, 1 is the foreground label and 0 is the background label.

(6c) The label l_biAs pixel point (x)_i,y_i) Tag S of₂(x_i,y_i) The S of₂(x_i,y_i) I.e. the image foreground to be extracted, where (x)_i,y_i)∈b_i。

8. The method of claim 7, wherein: super pixel b in step (6a)_iTag confidence Score of_biFor the super pixel B of the image B under the super pixel view angle_iExtracting result S of all contained pixels under pixel viewing angle₁The labels in (x, y) are summed, i.e.:

Score_bi＝∑l_ij

wherein, Score_biIs a super pixel b_iThe tag confidence of (1).