Finger vein feature fusion method based on cascade superpixel segmentation

Abstract

The present invention discloses a finger vein feature fusion method based on cascaded superpixel segmentation, including: 1) obtaining the ROI region of the finger vein image; 2) initializing k0 superpixel centers; 3) performing iterative superpixel segmentation on the ROI region to obtain a set P0 of large-scale image patches; 4) for each image patch in P0, initializing k1 superpixel centers; 5) performing iterative superpixel segmentation on each image patch in P0 to obtain a set P1 of medium-scale image patches; 6) for each image patch in P1, initializing k2 superpixel centers; 7) performing iterative superpixel segmentation on each image patch in P1 to obtain a set P2 of small-scale image patches; 8) extracting features for each image patch in P0, P1, and P2, concatenating the features of the same set to obtain a set feature vector, and then performing concatenation fusion of the feature vectors between different sets to obtain the fusion feature of the finger vein image. The present invention solves the problem in the prior art that the feature information expression ability is not strong enough due to single scale and unreasonable initialization of superpixel center points.

Description

Technical Field

[0001] The present invention relates to the technical field of finger vein feature extraction, and in particular to a finger vein feature fusion method based on cascade superpixel segmentation. Background Art

[0002] Finger vein recognition is a biometric recognition technology that collects the vein distribution image inside the finger to perform identity recognition. The current mainstream finger vein feature extraction is based on rectangular image blocks, but the rectangular grid-based block only considers the position information of the pixels, and does not consider the difference between vein patterns and non-vein patterns. Finger vein segmentation based on superpixels divides vein patterns and non-vein patterns into different clusters, better describes the local consistency of finger vein images, and achieves better results in feature extraction.

[0003] However, there are still some shortcomings in the finger vein feature extraction based on superpixel segmentation. When performing superpixel segmentation, the number of segmented image blocks needs to be manually selected, and the number of selected image blocks affects the performance of subsequent feature extraction; when initializing the superpixel center point, the pixel points evenly distributed on the finger vein image are used as the superpixel center. The uniform distribution cannot accurately reflect the distribution of the finger vein lines, resulting in the subsequent superpixel segmentation being unable to completely segment the vein lines and non-vein lines into different image blocks; only the local features of a single scale of the superpixel segmented image block are considered, and the feature information and global information of different scales are not considered. Therefore, how to improve the finger vein feature extraction based on superpixel segmentation is an urgent problem to be solved.

[0004] In summary, a finger vein feature fusion method based on cascade superpixel segmentation is invented, which has high practical application value. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, propose a finger vein feature fusion method based on cascaded superpixel segmentation, design a cascade structure of superpixel segmentation to obtain feature information of different scales from large-scale segmented image blocks to small-scale segmented image blocks, and propose a new superpixel center point initialization method to make the distribution of superpixel segmented image blocks more reasonable. Finally, the feature information of different scales is fused to realize the feature extraction of finger vein images, which solves the problem that the current finger vein SLIC segmentation algorithm has a single scale and unreasonable superpixel center point initialization, resulting in insufficient expression ability of the extracted feature information.

[0006] To achieve the above object, the technical solution provided by the present invention is: a finger vein feature fusion method based on cascade superpixel segmentation, comprising the following steps:

[0007] 1) Obtain a finger vein image, perform edge detection on the image, fit the midline and then perform rotation correction to obtain the ROI area, and perform size normalization and grayscale normalization on the ROI area to obtain the ROI image;

[0008] 2) Select k0 pixels on the ROI image obtained in step 1) as the initial k0 superpixel centers;

[0009] 3) According to the superpixel center initialized in step 2), the ROI image is iteratively segmented into superpixels using the SLIC algorithm to obtain a large-scale image block set P0 of superpixel segmentation;

[0010] 4) For each image block in the image block set P0 obtained in step 3), k1 pixels on the image block are selected as initialized k1 superpixel centers;

[0011] 5) According to the superpixel center initialized in step 4), use the SLIC algorithm to iteratively superpixel segment each image block in the image block set P0 to obtain a superpixel segmented mesoscale image block set P1;

[0012] 6) For each image block in the image block set P1 obtained in step 5), k2 pixels on the image block are selected as initialized k2 superpixel centers;

[0013] 7) According to the superpixel center initialized in step 6), use the SLIC algorithm to iteratively superpixel segment each image block in the image block set P1 to obtain a small-scale image block set P2 of superpixel segmentation;

[0014] 8) Extract features from each image block in P0, P1, and P2 respectively, connect the features of the same set in series to obtain a set feature vector, and then connect and fuse the feature vectors of different sets in series to obtain the fusion features of the finger vein image.

[0015] Further, in step 2), k0 pixels are selected as the initial k0 superpixel centers, and the specific steps are as follows:

[0016] 2.1) Take the upper left corner of the ROI image as the origin, the horizontal rightward as the x-axis, and the vertical downward as the y-axis to establish a coordinate system, randomly select a point in the ROI image as the superpixel center, and initialize i=1;

[0017] 2.2) The set of superpixel centers is S, and the sum of the distances D from each pixel a on the image to the superpixel center is calculated. a ;

[0018]

[0019]

[0020]

[0021]

[0022] Where j is a superpixel center point in the superpixel center point set, g j is the gray value of the superpixel center point, a is the pixel point on the image, and g a is the gray value of the pixel on the image, x j ,y j are the horizontal and vertical coordinates of the superpixel center, respectively. a ,y a are the horizontal and vertical coordinates of the pixel points, dg ja 、dc ja Represents the grayscale difference and position difference between a and j, N g 、N c They are grayscale normalization factor, position normalization factor, D ja represents the distance between a and j, D a Represents the sum of the distances from a to the superpixel center;

[0023] 2.3) Select D a The pixel corresponding to the maximum value is a new superpixel center point, added to the superpixel center set S, and i=i+1 is updated;

[0024] 2.4) Repeat steps 2.2) to 2.3) until i=k0.

[0025] Further, in step 3), the ROI image is iteratively segmented into superpixels using the SLIC algorithm according to the initialized superpixel center. The specific steps are as follows:

[0026] 3.1) Set the label label(a) of each pixel point a in the region to -1 and the distance dis(a) to +∞;

[0027] 3.2) For each pixel point a, calculate the distance D from the pixel point to the center point of each superpixel ja , if D ja <dis(a), then label(a)=j, dis(a)=D ja ;

[0028] 3.3) Update the coordinates of each superpixel center. The specific formula is as follows:

[0029]

[0030]

[0031] Where n is the number of label(a)=j pixels;

[0032] 3.4) Repeat steps 3.2) to 3.3) until the iteration number threshold I0 is reached to obtain the large-scale super-pixel segmentation image block P0.

[0033] Further, in step 4), for each image block in the acquired image block set P0, k1 superpixel centers are initialized, and the specific steps are as follows:

[0034] 4.1) For each image block of P0, randomly select a pixel point in the image block as the superpixel center point, and initialize the number i=1;

[0035] 4.2) Repeat steps 2.2) and 2.3) until i=k1.

[0036] Further, in step 5), the SLIC algorithm is used to iteratively superpixel segment each image block in the image block set P0 to obtain a superpixel segmented mesoscale image block set P1. The specific steps are as follows:

[0037] 5.1) Select an image block in P0, execute steps 3.1) to 3.3) until the iteration number threshold I1 is reached, obtain a superpixel segmented image block at the middle scale, and add it to the image block set P1;

[0038] 5.2) Repeat step 5.1) until the same operation is performed on each image block in P0.

[0039] Further, in step 6), for each image block in the acquired image block set P1, k2 superpixel centers are initialized, and the specific steps are as follows:

[0040] 6.1) For each image block of P1, randomly select a pixel point in the image block as the superpixel center point, and initialize the number i=1;

[0041] 6.2) Repeat steps 2.2) and 2.3) until i=k2.

[0042] Further, in step 7), the SLIC algorithm is used to iteratively perform superpixel segmentation on each image block in the image block set P1 to obtain a small-size image block set P2 of superpixel segmentation. The specific steps are as follows:

[0043] 7.1) For an image block in P1, execute steps 3.1) to 3.3) until the iteration number threshold I2 is reached, and a super-pixel segmented image block at a small scale is obtained, and added to the image block set P2;

[0044] 7.2) Repeat step 7.1) until the same operation is performed on each image block in P1.

[0045] Furthermore, in step 8), LBP texture features and three different scalar features, namely, average LBP value, average gray value, and image entropy, are extracted from each image block in P0, P1, and P2 respectively. The features of the same set are connected in series to obtain a set feature vector, and then the feature vectors of different sets are connected in series and fused to obtain the fusion features of the finger vein image.

[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0047] 1. The finger vein feature fusion method of the present invention designs a cascade structure of superpixel segmentation, obtains feature information of different scales from large-scale segmented image blocks to small-scale segmented image blocks, and fuses feature information of different scales.

[0048] 2. Compared with the single-scale superpixel segmentation method, the number of finger vein image segmentation blocks obtained by cascade superpixel segmentation does not increase significantly, thereby ensuring that the feature dimension does not increase significantly, that is, it has a stronger ability to express feature information while the algorithm complexity remains basically unchanged.

[0049] 3. Obtaining feature information at different scales can alleviate the impact of the number of superpixel center points on feature extraction at a single scale to a certain extent.

[0050] 4. A new superpixel center point initialization method is proposed, which can more accurately reflect the distribution of finger vein patterns, thereby making the superpixel segmentation image blocks more reasonable. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Flow chart of the method of the present invention.

[0052] Figure 2 Flowchart for initializing superpixel centers.

[0053] Figure 3 Schematic diagram of the ROI image after finger vein preprocessing.

[0054] Figure 4 This is a schematic diagram of a large-scale image block set obtained after the first step of segmentation of the ROI image. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0056] like Figures 1 to 4 As shown, this embodiment provides a finger vein feature fusion method based on cascade superpixel segmentation, comprising the following steps:

[0057] 1) Obtain a finger vein image, perform edge detection on the image, fit the midline and then perform rotation correction to obtain the ROI area, perform scale normalization and grayscale normalization on the ROI area to obtain the ROI image, such as Figure 3 shown.

[0058] 2) Initialize 7 superpixel centers of the ROI image. The initialization process is as follows: Figure 2 As shown, the specific steps of initialization are as follows:

[0059] 2.1) Establish a coordinate system with the upper left corner of the ROI area as the origin, the horizontal rightward as the x-axis, and the vertical downward as the y-axis. Randomly select a point in the ROI area as the superpixel center, and initialize i=1;

[0060] 2.2) The set of superpixel centers is S, and the sum of the distances from each pixel a in the region to the superpixel center is calculated as D a ;

[0061]

[0062]

[0063]

[0064]

[0065] Where j is a superpixel center point in the superpixel center point set, g j is the gray value of the center point of the superpixel in the superpixel, a is the pixel point in the area, g a is the gray value of the pixel in the area, x j ,y j are the horizontal and vertical coordinates of the superpixel center, respectively. a ,y a are the horizontal and vertical coordinates of the pixel points, dg ja 、dc ja Represents the grayscale difference and position difference between a and j, N g 、N c They are grayscale normalization factor and position normalization factor respectively, and N is set g =1, D ja represents the distance between a and j, D a Represents the sum of the distances from a to the superpixel center.

[0066] 2.3) Select D a The pixel corresponding to the maximum value is a new superpixel center point, added to the superpixel center set S, and i=i+1 is updated;

[0067] 2.4) Repeat steps 2.2) to 2.3) until i=7.

[0068] 3) According to the 7 superpixel center points initialized on the ROI image in step 2), the ROI region is iteratively segmented into superpixels using the SLIC segmentation algorithm to obtain a large-scale image block set P0 of superpixel segmentation. The specific steps are as follows:

[0069] 3.1) Set the label label(a) of each pixel point a in the region to -1 and the distance dis(a) to +∞;

[0070] 3.2) For each pixel point a, calculate the distance D from the pixel point to the center point of each superpixel ja , if D ja <dis(a), then label(a)=j, dis(a)=D ja ;

[0071] 3.3) Update the coordinates of each superpixel center. The specific formula is as follows:

[0072]

[0073]

[0074] Where n is the number of label(a)=j pixels;

[0075] 3.4) Repeat steps 3.2) to 3.3) until the iteration threshold I0 is reached, set I0 = 10, and obtain the large-scale image block set P0 of superpixel segmentation, as shown in Figure 4 shown.

[0076] 4) For each image block in the acquired image block set P0, initialize three superpixel centers. The specific steps are as follows:

[0077] 4.1) Randomly select a point in the P0 image block as the superpixel center and initialize i=1;

[0078] 4.2) Repeat steps 2.2) and 2.3) until i=3.

[0079] 5) Use the SLIC algorithm to iteratively superpixel segment each image block in the image block set P0 to obtain a superpixel segmented medium-scale image block set P1. The specific steps are as follows:

[0080] 5.1) For an image block in P0, execute steps 3.1) to 3.3) until the iteration threshold I1 is reached, set I1 = 10, obtain the superpixel segmented image block at the mesoscale, and add it to the mesoscale image block set P1;

[0081] 5.2) Repeat 5.1) and perform the same operation on each image block in P0.

[0082] 6) For each image block in the acquired image block set P1, initialize two superpixel centers. The specific steps are as follows:

[0083] 6.1) Randomly select a point in the P1 image block as the superpixel center and initialize i=1;

[0084] 6.2) Repeat steps 2.2) and 2.3) until i=2.

[0085] 7) Use the SLIC algorithm to iteratively perform superpixel segmentation on each image block in the image block set P1 to obtain a small-scale image block set P2 of superpixel segmentation. The specific steps are as follows:

[0086] 7.1) For an image block in P1, execute steps 3.1) to 3.3) until the iteration number threshold I2 is reached, set I2 = 10, obtain a superpixel segmented image block at a small scale, and add it to the small scale image block set P2;

[0087] 7.2) Repeat 7.1) and perform the same operation on each image block in P1.

[0088] 8) LBP texture features and three different scalar features, namely, average LBP value, average gray value, and image entropy, are extracted from each image block in P0, P1, and P2 respectively. The features of the same set are connected in series to obtain the set feature vector, and then the feature vectors of different sets are connected in series to obtain the fusion features of the finger vein image. The experimental process ends.

[0089] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. Finger vein feature fusion method based on cascade superpixel segmentation, characterized in that: The following steps are involved: 1) Obtain the finger vein image, perform edge detection on the image, fit the midline and perform rotation correction to obtain the ROI area, and perform size normalization and grayscale normalization on the ROI area to obtain the ROI image; 2) Select the ROI image obtained in step 1) Pixels are used as initialization The specific steps are as follows: 2.1) Take the upper left corner of the ROI image as the coordinate origin, and Axis, vertical downward Axis, establish the coordinate system, randomly select a point in the ROI image as the superpixel center, and initialize ; 2.2) The set of superpixel center points is , calculate each pixel on the image The sum of the distances to the superpixel center ; ； ； ； ； In the formula, is a superpixel center point in the superpixel center point set, is the gray value of the superpixel center point, is the pixel point on the image, is the gray value of the pixel on the image, , are the horizontal and vertical coordinates of the superpixel center point, , are the horizontal and vertical coordinates of the pixel points, , Respectively represent and The grayscale difference and position difference of , They are grayscale normalization factor and position normalization factor respectively. represent and The distance represent The sum of distances to superpixel centers; 2.3) Selection The pixel corresponding to the maximum value is a new superpixel center point and is added to the superpixel center set. ,renew ; 2.4) Repeat steps 2.2) to 2.3) until ; 3) According to the superpixel center initialized in step 2), the ROI image is iteratively segmented using the SLIC algorithm to obtain a large-scale image block set of superpixel segmentation , the specific steps are as follows: 3.1) Each pixel in the area The label label(a) is set to -1 and the distance dis(a) is set to ; 3.2) For each pixel , calculate the distance from the pixel to the center of each superpixel ,if ,but , ; 3.3) Update the coordinates of each superpixel center. The specific formula is as follows: ； ； Where n is The number of pixels; 3.4) Repeat steps 3.2) to 3.3) until the iteration threshold is reached. , get the super-pixel segmentation image block at a large scale ; 4) For the image block set obtained in step 3) For each image block in Pixels are used as initialization The specific steps are as follows: 4.1) For For each image block, a pixel point of the image block is randomly selected as the superpixel center point, and the number of initialization ; 4.2) Repeat steps 2.2) and 2.3) until ; 5) Based on the superpixel centers initialized in step 4), use the SLIC algorithm to Iterate superpixel segmentation for each image block in the image to obtain a set of medium-scale image blocks of superpixel segmentation ; 6) For the image block set obtained in step 5) For each image block in Pixels are used as initialization The specific steps are as follows: 6.1) For For each image block, a pixel point of the image block is randomly selected as the superpixel center point, and the number of initialization ; 6.2) Repeat steps 2.2) and 2.3) until ; 7) Based on the superpixel centers initialized in step 6), use the SLIC algorithm to Iterate superpixel segmentation for each image block in the image to obtain a set of small-scale image blocks segmented by superpixel ; 8) Respectively , , Features are extracted from each image block in the image, and the features of the same set are connected in series to obtain a set feature vector. The feature vectors of different sets are then connected in series and fused to obtain the fusion features of the finger vein image.

2. The finger vein feature fusion method based on cascade superpixel segmentation according to claim 1 is characterized in that: In step 5), the SLIC algorithm is used to perform Iterate superpixel segmentation for each image block in the image to obtain a set of medium-scale image blocks of superpixel segmentation , the specific steps are as follows: 5.1) Selection For an image block in , execute steps 3.1) to 3.3) until the iteration number threshold is reached. , get the super-pixel segmentation image block at the medium scale and add it to the image block set ; 5.2) Repeat step 5.1) until The same operation is performed on each image block in .

3. The finger vein feature fusion method based on cascade superpixel segmentation according to claim 2 is characterized in that: In step 7), the SLIC algorithm is used to perform Iterate superpixel segmentation for each image block in the image to obtain a set of small-sized image blocks segmented by superpixel , the specific steps are as follows: 7.1) For For an image block in , execute steps 3.1) to 3.3) until the iteration number threshold is reached. , get the super-pixel segmentation image block at a small scale and add it to the image block set ; 7.2) Repeat step 7.1) until The same operation is performed on each image block in .

4. The finger vein feature fusion method based on cascade superpixel segmentation according to claim 3 is characterized in that: In step 8), , , LBP texture features and three different scalar features, namely, average LBP value, average gray value, and image entropy, are extracted from each image block. The features of the same set are connected in series to obtain the set feature vector, and then the feature vectors of different sets are connected in series to obtain the fusion features of the finger vein image.

Images