CN110648341B - Target boundary detection method based on scale space and subgraph - Google Patents

Abstract

The invention relates to a target boundary detection method based on a scale space and a subgraph. The method comprises the following steps: step 1, inputting a digital image, step 2, setting a mask window set, step 3, constructing a scale space, step 4, decomposing the scale space, step 5, extracting point feature vectors, step 6, calculating a region estimation value, step 7, extracting a candidate target region set, step 8, dividing the candidate target region set, step 9, calculating the reliability of the divided region set, step 10, superpixel division, step 11, calculating the average reliability of the superpixel block set, step 12, nonlinear transformation processing, step 13, binarization and boundary extraction, and step 14, outputting a target boundary. The method solves the problem of low accuracy of target boundary detection in the digital image, and achieves the effect of full automation. The method can be used in the fields of ground feature extraction of remote sensing images, target recognition of robot vision and the like.

Description

An object boundary detection method based on scale space and subgraph

technical field

The invention relates to the field of digital image processing, in particular to a target boundary detection method based on scale space and subgraph.

Background technique

Object boundary detection is a fundamental problem in image processing and computer vision, and its main purpose is to identify locations in digital images with significant brightness changes. Significant changes in image properties usually reflect important events and changes in properties, these include discontinuities in depth, discontinuities in surface orientation, changes in material properties, and changes in scene lighting, among others. Object boundary detection is a research field in image processing and computer vision, especially feature extraction. With the continuous development of science and technology, some new technologies and methods have emerged in the field of target boundary detection in recent years, such as wavelet multi-scale method, fractal theory method, mathematical morphology method, artificial intelligence and genetic algorithm, etc. It has strong noise performance and can better detect the boundary of the target. However, the current target boundary detection methods can only achieve good results within a certain range of conditions, and each method has certain shortcomings, especially the balance of detection accuracy and anti-noise ability. Some methods have high boundary detection accuracy, but poor anti-noise performance; some methods have strong anti-noise performance, but the detection accuracy is not enough; although some methods can solve the coordination problem of the above two to a certain extent, However, the algorithm is complex and the operation time is long, and there is no boundary detection method suitable for various requirements. Therefore, it is always an important direction of image processing and analysis research to find a target boundary detection algorithm with simple algorithm, which can better solve the coordination problem of boundary detection accuracy and anti-noise performance.

SUMMARY OF THE INVENTION

The present invention provides a target boundary detection method based on scale space and sub-image, which can overcome the difficulty of detection of target boundary in current digital images. By using the scale space and sub-image sets as masks, various targets in digital images can be detected without the need for Manual intervention, high degree of automation.

The technical scheme adopted for realizing the object of the present invention is:

Step 1: Input digital image I;

Step 2: set a mask window set SW in the digital image I, and the size of each window included in the mask window set SW is L ₁ ×L ₂ ;

Step 3: Construct scale space SI based on digital image I;

Step 4: Use the mask window set SW to decompose all the images contained in the scale space SI to obtain the subgraph set SR;

Step 5: Extract the point feature gf _xy of the pixel (x, y) in the sub-image SR _a included in the sub-image set SR, and form the point feature vector GF _a of the sub-image SR _a by the point features of all the pixels in the sub-image SR _a , where , 1≤a≤A, A is the number of subgraphs in the subgraph set SR, GF _a =(gf ₁ ,gf ₂ ,gf ₃ ,…,gf _(y-1)×L+x ,…,gf _{L1× L2} ), where 1≤x≤L ₁ , 1≤y≤L ₂ , that is, GF _a is a L ₁ ×L ₂ -dimensional row vector;

Step 6: Calculate the area estimate V _SRa of the subgraph SR _a by the following formula:

V _SRa =GF _a ·WT

In the formula, the symbol · represents the inner product operation, WT is the template weight vector of the subgraph, and its vector dimension is consistent with GF _a ;

Step 7: Set the threshold V _T , when V _SRa is greater than V _T , keep the sub-image SR _a as the candidate target area, otherwise, delete the sub-image SR _a ;

Step 8: Iteratively execute steps 5-7 until all subgraphs SR _a included in the subgraph set SR are traversed, and the candidate target area set SC is obtained;

Step 9: Divide the area SC _b contained in the candidate target area set SC to obtain the divided area set SA contained in the area SC _b , where 1≤b≤B, and B is the number of areas in the candidate target area set SC;

Step 10: Calculate the credibility CV _c of the segmentation area SA _c included in the segmentation area set SA by the following formula, and assign the credibility CV _c of the segmentation area SA _c to each pixel included in the segmentation area SA _c , wherein, 1≤c≤C, C is the number of elements in the segmented area set SA,

In the formula, SM(SA _c , SA _q ) is the grayscale similarity between the segmented area SA _c and the segmented area SA _q , D _s (SA _c , SA _q ) is the difference between the segmented area SA _c and the segmented area SA _q The adjacent distance is defined as the Euclidean distance between the centroids of the segmented area SA _c and the segmented area SA _q , α is the contribution control coefficient of the adjacent distance D _s (SA _c , SA _q ), the smaller the value, the closer The greater the contribution of the distance D _s (SA _c , SA _q ) to the reliability CV _c , w _c is the ratio of the number of pixels in the segmented area SA _c to the number of pixels in the area SC _b , which is used to measure the effectiveness of the segmented area SA _c . size;

Step 11: Iteratively execute steps 9-10 until traversing all the segmented regions SA _c included in the segmented region set SA and traversing all the regions SC _b included in the candidate target region set SC;

Step 12: carry out superpixel segmentation to input digital image I, obtain superpixel block set SSB;

Step 13: iteratively calculate the average reliability ACV _d of the superpixel block SSB _d included in the superpixel block set SSB, wherein 1≤d≤D, and D is the number of superpixel blocks;

Step 14: Perform nonlinear transformation processing on the average reliability ACV _d of the superpixel block SSB _d by the following formula to obtain the transformed average reliability EACV _{d of the superpixel block SSB d :}

The transformed average reliability EACV _d of the superpixel block SSB _d is assigned to each pixel included in the superpixel block SSB _d .

Step 15: iteratively execute steps 13-14, until traversing all the superpixel blocks SSB _d included in the superpixel block set SSB, obtain the transformed image Ic that is formed of the pixel whose value is the transformed average reliability;

Step 16: carry out binarization processing to the transformed image Ic, and carry out boundary extraction to the binarization result to obtain a boundary set S _edge ;

Step 17: Superimpose the boundary set S _edge on the original image as the output result. The enhancement method described in step 2 adopts the histogram equalization method.

The positions of the mask window set SW described in step 2 are random, each window can overlap, and the mask window set SW completely covers the digital image I.

The scale space SI described in step 3 includes images I,I ₁ ,I ₂ ,...,In , where _n is the number of layers of the scale space SI,

...

为卷积运算符，g(x,y,t)为二维卷积核，

(x,y)为图像中的像素坐标，t为尺度空间因子。In the formula,

is the convolution operator, g(x, y, t) is the two-dimensional convolution kernel,

(x, y) are the pixel coordinates in the image, and t is the scale space factor.

The method for decomposing the scale space SI described in step 4 is: using the mask window SW _h included in the mask window set SW to perform AND operation with all the images included in the scale space SI respectively, and intercepting the image area corresponding to the mask window SW _h As a subgraph, where 1≤h≤H, H is the number of subgraphs included in the subgraph set SR, and the above steps are iteratively executed until all the mask windows SW _h included in the mask window set SW are traversed.

The extraction method of the point feature gf _xy of the pixel (x, y) described in step 5 is:

gf _xy = min(|G _H (x,y)|+|G _V (x,y)|,255)

In the formula, G _H (x,y)=I(x+1,y)-I(x-1,y), G _V (x,y)=I(x,y+1)-I(x, y-1), where |G _H (x, y)| is the gradient modulus value of the pixel (x, y) in the horizontal direction, |G _V (x, y)| is the pixel (x, y) in the vertical direction The gradient modulus value of , min() is the minimum function.

The grayscale similarity SM (SA _c , SA _q ) between the segmented area SA _c and the segmented area SA _q described in step 10 is calculated by the following formula:

为灰度值n1在分割区域SA_c中出现的概率，

为灰度值n2在分割区域SA_q中出现的概率，O(n1,n2)为灰度值n1和灰度值n2之间的欧氏距离。In the formula, n1 and n2 are the grayscale values of the segmented area SA _c and the segmented area SA _q , respectively, N1 and N2 are the gray levels of the segmented area SA _c and the segmented area SA _q , respectively,

is the probability that the gray value n1 appears in the segmented area SA _c ,

is the probability that the gray value n2 appears in the segmented area SA _q , and O(n1, n2) is the Euclidean distance between the gray value n1 and the gray value n2.

The beneficial effects of the invention are as follows: the problem that the detection accuracy of the target boundary in the digital image is not high is solved, and the effect of complete automation is achieved. It can be used in the fields of object extraction of remote sensing images and target recognition of robot vision.

Description of drawings

Fig. 1 is the overall processing flow chart of the present invention;

FIG. 2 is a schematic diagram of the mask window arrangement of the present invention.

Detailed ways

The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

FIG. 1 is a flow chart of the overall processing of the present invention. As shown in Figure 1, 101 is the step of inputting a digital image, 102 is the step of setting the mask window set, 103 is the step of constructing the scale space, 104 is the step of decomposing the scale space, 105 is the step of extracting point feature vectors, and 106 is the step of calculating the region estimate Step, 107 is the step of extracting the candidate target area set, 108 is the step of dividing the candidate target area set, 109 is the step of calculating the reliability of the set of dividing areas, 110 is the step of superpixel segmentation, 111 is the average probability of calculating the set of superpixel blocks. In the reliability step, 112 is the nonlinear transformation processing step, 113 is the binarization and boundary extraction step, and 114 is the output target boundary step.

Step 101: input a grayscale digital image I with a grayscale level of 256;

Step 102: Set a mask window set SW in the digital image I, the size of each window included in the mask window set SW is L ₁ ×L ₂ , and set L ₁ =L ₂ =40, wherein the mask window set The position of SW is random. In order to improve the accuracy of the target, each window can overlap, and the mask window set SW completely covers the digital image I;

Step 103: Construct a scale space SI based on the digital image I, and the scale space _SI includes images I, I ₁ , I ₂ , . n is set to 4,

为卷积运算符，g(x,y,t)为二维卷积核，

(x,y)为图像中的像素坐标，t为尺度空间因子；In the formula,

is the convolution operator, g(x, y, t) is the two-dimensional convolution kernel,

(x, y) is the pixel coordinates in the image, and t is the scale space factor;

Step 104: Use the mask window set SW to decompose all images contained in the scale space SI to obtain a subgraph set SR. The specific method for decomposing the scale space SI is: use the mask window SW _h contained in the mask window set SW to separate the scale space Perform AND operation on all images included in SI, and intercept the image area corresponding to the mask window SW _h as a sub-image, where 1≤h≤H, H is the number of sub-images included in the sub-image set SR, and iteratively execute the above steps until Traverse all mask windows SW _h contained in the mask window set SW;

Step 105: Extract the point feature gf _xy of the pixel (x, y) in the sub-image SR _a included in the sub-image set SR, and form the point feature vector GF _a of the sub-image SR _a by the point features of all the pixels in the sub-image SR _a , where , 1≤a≤A, A is the number of subgraphs in the subgraph set SR, GF _a =(gf ₁ ,gf ₂ ,gf ₃ ,…,gf _(y-1)×L+x ,…,gf _{L1× L2} ), where 1≤x≤L ₁ , 1≤y≤L ₂ , since L ₁ =L ₂ =40, GF _a is a 40×40=1600-dimensional row vector, the point of pixel (x, y) The extraction method of feature gf _xy is:

gf _xy = min(|G _H (x,y)|+|G _V (x,y)|,255)

In the formula, G _H (x,y)=I(x+1,y)-I(x-1,y), G _V (x,y)=I(x,y+1)-I(x, y-1), where |G _H (x, y)| is the gradient modulus value of the pixel (x, y) in the horizontal direction, |G _V (x, y)| is the pixel (x, y) in the vertical direction The gradient modulus value of , min() is the minimum function;

Step 106: Calculate the region estimated value V _SRa of the sub-graph SR _a by the following formula:

V _SRa =GF _a ·WT

The region estimation value V _SR is used to evaluate the possibility that the sub-graph SR _a contains the target. The larger the value, the greater the possibility of containing the target. In the formula, the symbol · represents the inner product operation, WT is the template weight vector of the subgraph, and its vector dimension is consistent with GF _a , which is 1600 dimensions;

Step 107: Set a threshold value V _T , when V _SRa is greater than V _T , keep the sub-image SR _a as a candidate target area, otherwise, delete the sub-image SR _a ;

Steps 105-107 are iteratively executed until all subgraphs SR _a contained in the subgraph set SR are traversed to obtain the candidate target area set SC;

Step 108: Use the region growing segmentation algorithm to segment the region SC _b included in the candidate target region set SC to obtain the segmented region set SA included in the region SC _b , where 1≤b≤B, and B is the the number of areas;

Step 109: Calculate the reliability CV _c of the segmented area SA _c included in the segmented area set SA by the following formula, and assign the reliability CV _c of the segmented area SA _c to each pixel included in the segmented area SA _c , wherein, 1≤c≤C, C is the number of elements in the segmented area set SA,

In the formula, SM(SA _c , SA _q ) is the grayscale similarity between the segmented area SA _c and the segmented area SA _q , D _s (SA _c , SA _q ) is the difference between the segmented area SA _c and the segmented area SA _q The adjacent distance is defined as the Euclidean distance between the centroids of the segmented area SA _c and the segmented area SA _q , α is the contribution control coefficient of the adjacent distance D _s (SA _c , SA _q ), the smaller the value, the closer The greater the contribution of the distance D _s (SA _c , SA _q ) to the reliability CV _c , w _c is the ratio of the number of pixels in the segmented area SA _c to the number of pixels in the area SC _b , which is used to measure the effectiveness of the segmented area SA _c . size;

The grayscale similarity SM(SA _c , SA _q ) between the segmented area SA _c and the segmented area SA _q is calculated by the following formula:

为灰度值n1在分割区域SA_c中出现的概率，

为灰度值n2在分割区域SA_q中出现的概率，O(n1,n2)为灰度值n1和灰度值n2之间的欧氏距离；In the formula, n1 and n2 are the grayscale values of the divided area SA _c and the divided area SA _q , respectively, N1 and N2 are the gray level of the divided area SA _c and the divided area SA _q , respectively, it can be known from step 101: N1=N2= 256,

is the probability that the gray value n1 appears in the segmented area SA _c ,

is the probability that the gray value n2 appears in the segmented area SA _q , O(n1, n2) is the Euclidean distance between the gray value n1 and the gray value n2;

Steps 108-109 are iteratively executed until all the segmented regions SA _c included in the segmented region set SA and all regions SC _b included in the candidate target region set SC are traversed;

Step 110: carry out superpixel segmentation to input digital image I, obtain superpixel block set SSB;

Step 111: Iteratively calculate the average reliability ACV _d of the superpixel block SSB _d included in the superpixel block set SSB, wherein 1≤d≤D, and D is the number of superpixel blocks;

Step 112: Perform nonlinear transformation processing on the average reliability ACV _d of the superpixel block SSB _d by the following formula to obtain the transformed average reliability EACV _{d of the superpixel block SSB d :}

And the transformation average reliability EACV _d of super pixel block SSB _d is assigned to each pixel that super pixel block SSB _d comprises

Iteratively execute steps 111-112, until traversing all superpixel blocks SSB _d included in the superpixel block set SSB, obtain the transformed image Ic that is formed by taking the value of the pixel of the transformed average reliability;

Step 113: Binarization processing is performed on the transformed image Ic, and boundary extraction is performed on the binarization result to obtain a boundary set S _edge ;

Step 114: Superimpose the boundary set S _edge on the original image as an output result.

FIG. 2 is a schematic diagram of the mask window arrangement of the present invention. As shown in Figure 2, 201 is the input image I, 202 is a mask window included in the mask window set SW, 203 is the overlapping part between the two mask windows, the input image captured by the mask window (202) The region of I(201) is a subgraph in the subgraph set SR.

Claims (6)

1. A target boundary detection method based on scale space and subgraph is characterized by comprising the following steps:

step 1: inputting a digital image I;

step 2: setting a mask window set SW in the digital image I, wherein the size of each window contained in the mask window set SW is L ₁ ×L ₂ ；

And 3, step 3: constructing a scale space SI based on the digital image I;

and 4, step 4: decomposing all images contained in the scale space SI by using a mask window set SW to obtain a sub-image set SR;

and 5: extracting subgraph SR contained in subgraph set SR _a Dot characteristic gf of middle pixel (x, y) _xy From the sub-diagram SR _a The point characteristics of all the pixels in the sub-graph SR _a Point feature vector GF of _a Wherein a is more than or equal to 1 and less than or equal to A, A is the number of subgraphs in the subgraph set SR, GF _a ＝(gf ₁ ,gf ₂ ,gf ₃ ,…,gf _(y-1)×L+x ,…,gf _L1×L2 ) Wherein x is more than or equal to 1 and less than or equal to L ₁ ，1≤y≤L ₂ I.e. GF _a Is L ₁ ×L ₂ A row vector of dimensions;

step 6: calculation of the subgraph SR by _a Is estimated by the region of _SRa ：

V _SRa ＝GF _a ·WT

Where the symbol represents an inner product operation, WT is the template weight vector of the subgraph, its vector dimension and GF _a The consistency is achieved;

and 7: setting a threshold value V _T When V is _SRa Greater than V _T When, sub-graph SR _a Reserved as a candidate target region, otherwise, the sub-graph SR is deleted _a ；

And 8: iteratively executing steps 5-7 until all sub-graphs SR included in the sub-graph set SR are traversed _a Obtaining a candidate target area set SC;

and step 9: for the region SC contained in the candidate target region set SC _b Dividing to obtain regions SC _b B is more than or equal to 1 and less than or equal to B, wherein B is the number of the areas in the candidate target area set SC;

step 10: the divided regions SA included in the divided region set SA are calculated by the following formula _c Reliability CV of (a) _c And dividing the area SA _c Reliability CV of _c Assign to a partitioned area SA _c Each pixel is contained, wherein C is more than or equal to 1 and less than or equal to C, C is the number of elements in the set SA of the divided areas,

in the formula, SM (SA) _c ,SA _q ) As a divided area SA _c And a divided area SA _q Gray scale similarity therebetween, D _s (SA _c ,SA _q ) As a divided area SA _c And a divided area SA _q The adjacent distance between them is defined as the dividing area SA _c And a divided area SA _q The Euclidean distance between two centroids, alpha being the proximity distance D _s (SA _c ,SA _q ) The smaller the value of the contribution degree control coefficient is, the closer the distance D is _s (SA _c ,SA _q ) For reliability CV _c The greater the contribution of (a), w _c As a divided area SA _c Occupied area SC of the number of pixels _b Is used for measuring the division area SA _c The size of (d);

step 11: the steps 9-10 are iteratively executed until all the segmented areas SA comprised by the segmented area set SA are traversed _c And traversing all the regions SC contained in the candidate target region set SC _b ；

Step 12: performing superpixel segmentation on an input digital image I to obtain a superpixel block set SSB;

step 13: iteratively calculating the super-pixel block SSB included in the super-pixel block set SSB _d Average reliability of ACV _d D is more than or equal to 1 and less than or equal to D, and D is the number of the super pixel blocks;

step 14: for super pixel block SSB by _d Average reliability of ACV _d Performing nonlinear transformation to obtain super pixel block SSB _d Mean likelihood of transformation of EACV _d ：

And the super pixel block SSB _d Mean likelihood of transformation of EACV _d Assign to super pixel block SSB _d Each pixel contained;

step 15: steps 13-14 are iteratively performed until all super pixel blocks SSB included in the super pixel block set SSB are traversed _d Obtaining a transformation image Ic consisting of pixels with the values of transformation average reliability;

step 16: binarization processing is carried out on the converted image Ic, and boundary extraction is carried out on the binarization result to obtain a boundary set S _edge ；

And step 17: set the boundary S _edge And the image is superposed on the original image as an output result.

2. The method of claim 1, wherein the mask window set SW of step 2 is randomly positioned, each window can overlap, and the mask window set SW completely covers the digital image I.

3. The method as claimed in claim 1, wherein the scale space SI in step 3 comprises an image I, I ₁ ,I ₂ ,…,I _n Wherein n is the number of layers of the scale space SI,

...

in the formula (I), the compound is shown in the specification,

for the convolution operator, g (x, y, t) is a two-dimensional convolution kernel,

(x, y) are the pixel coordinates in the image, and t is the scale space factor.

4. The method of claim 1, wherein the method of decomposing the scale space SI in step 4 comprises: mask window SW contained by mask window set SW _h Respectively AND-calculating with all images contained in the scale space SI, intercepting the mask window SW _h Taking the image area at the corresponding position as a subgraph, wherein H is more than or equal to 1 and less than or equal to H, H is the number of subgraphs contained in the subgraph set SR, and iteratively executing the steps until all mask windows SW contained in the mask window set SW are traversed _h 。

5. The method of claim 1, wherein the step 5 comprises a step of detecting a point feature gf of the pixel (x, y) _xy The extraction method comprises the following steps:

gf _xy ＝min(|G _H (x,y)|+|G _V (x,y)|,255)

in the formula, G _H (x,y)＝I(x+1,y)-I(x-1,y)，G _V (x, y) ═ I (x, y +1) -I (x, y-1), where, | G _H (x, y) | is a gradient modulus of the pixel (x, y) in the horizontal direction, | G _V (x, y) | is a gradient modulus value of the pixel (x, y) in the vertical direction, and min () is a minimum function.

6. The method of claim 1, wherein the segmented region SA of step 10 is a scale space and sub-graph-based object boundary detection method _c And a divided area SA _q Similarity of gray scale SM (SA) _c ,SA _q ) Calculated by the following formula:

wherein n1 and n2 denote divisional areas SA _c And a divided area SA _q N1 and N2 are respectivelyPartitioned area SA _c And a divided area SA _q Gray level of p _SAc (n1) Gray value n1 in the divided region SA _c Probability of occurrence of p _SAq (n2) Gray value n2 in the divided region SA _q O (n1, n2) is the euclidean distance between the gradation value n1 and the gradation value n 2.

Images