WO2020083407A1

WO2020083407A1 - Three-dimensional finger vein feature extraction method and matching method therefor

Info

Publication number: WO2020083407A1
Application number: PCT/CN2019/113883
Authority: WO
Inventors: 康文雄; 龚启琛
Original assignee: 华南理工大学
Priority date: 2018-10-23
Filing date: 2019-10-29
Publication date: 2020-04-30
Also published as: CN109543535A; AU2019368520A1; CN109543535B; AU2019368520B2

Abstract

A three-dimensional finger vein feature extraction method, comprising: step one: a two-dimensional finger vein image is acquired; step two: the two-dimensional finger vein image is mapped onto a three-dimensional model to construct a three-dimensional finger model; step three: the three-dimensional finger model is normalized; and step four: feature extraction is performed on the normalized three-dimensional finger model. A three-dimensional finger vein feature matching method. In said matching method, vein pattern feature and central axis geometric distance feature scores are calculated for a template sample and a sample to be matched, weighted fusion is performed, and determination is performed on fused matching scores by means of a threshold, thereby completing three-dimensional finger vein matching and identification. By means of the present three-dimensional finger vein feature extraction method and the matching method therefor, more vein pattern features can be extracted, resulting in better matching and identification results, and the problem of poor matching and identification performance caused by changes in finger position can be effectively solved, thereby improving the accuracy and effectiveness of vein matching and recognition.

Description

Three-dimensional finger vein feature extraction method and matching method

Technical field

The present invention relates to the technical field of vein recognition, and more specifically, to a three-dimensional finger vein feature extraction method and matching method.

Background technique

Biometric recognition technology is a technology that uses one or more human physiological characteristics (such as fingerprints, faces, irises, veins, etc.) or behavioral characteristics (such as gait, signature, etc.) for identity authentication. Among them, the finger vein recognition technology begins to occupy an important position in the field of identity authentication with its unique advantages. It is a biometric recognition technology that uses the vein information of the blood vessels under the skin of the finger as individual identity verification. In biometrics recognition, finger vein recognition technology has unique application advantages and application prospects due to its high security and strong stability. At present, finger vein recognition systems are based on two-dimensional vein images for recognition. When the finger is placed improperly, especially when the finger is rotated axially, its recognition performance will be greatly reduced.

However, there are relatively few studies on finger vein recognition in different postures. A few existing studies include: using ellipse model to expand the collected two-dimensional finger vein images, so as to standardize the finger vein images, and then intercept the effective area for Matching; using the circular model to expand the two-dimensional finger vein image; or, using the three-dimensional model method, the key is still to use the ellipse model to standardize the finger vein images in six different poses before matching. Regardless of which physical model is used, the situation where there are large differences between vein images taken by the same finger in different poses is improved to a certain extent, but the problem still remains: on the one hand, the corresponding texture area becomes less However, it is not conducive to matching; on the other hand, the quality of the vein image in the edge area is generally affected by the imaging factors, which also affects the recognition result. There is another method based on the multi-view geometry 3D imaging method, but this method is difficult to find or even find matching feature points during 3D reconstruction, so it is difficult to calculate the depth information of all vein textures. In addition, this method collects The vein texture is also only one-sided, so there is still a problem of limited feature information.

Summary of the invention

The purpose of the present invention is to overcome the shortcomings and deficiencies in the prior art, to provide a three-dimensional finger vein feature extraction method and matching method, the method can obtain more vein texture features, get a better matching recognition effect, and at the same time It can effectively solve the problem of poor matching recognition performance caused by the change of finger posture, thereby improving the accuracy and effectiveness of vein matching recognition.

In order to achieve the above object, the present invention is implemented by the following technical solution: A three-dimensional finger vein feature extraction method, characterized in that it includes the following steps:

In the first step, three fingers are used to capture the finger vein from three angles at equal angles to obtain a two-dimensional finger vein image;

In the second step, by calculating the parameters of the three cameras, the two-dimensional finger vein image is mapped to the three-dimensional model to realize the construction of the three-dimensional finger model;

The third step is to normalize the three-dimensional finger model to achieve the effect of eliminating the horizontal and vertical offset of the finger;

In the fourth step, feature extraction is performed on the normalized 3D finger model:

(1) Process the normalized three-dimensional finger model to generate three-dimensional texture expansion map and geometric distance feature map;

(2) The convolutional neural network is used to extract features from the three-dimensional texture expansion map and the geometric distance feature map to obtain vein texture features and the central axis geometric distance features; at the same time, the neural network is trained.

In the second step, by calculating the parameters of the three cameras, the two-dimensional finger vein image is mapped to the three-dimensional model, and the construction of the three-dimensional finger model means that the finger profile is approximately regarded as an ellipse, and the three-dimensional finger is divided into equal distances into For several profiles, calculate the profile of each profile, and use multiple ellipses with different radii and different positions to approximate the model of the finger; then connect all the profiles in series according to the direction of the central axis of the finger to obtain an approximate three-dimensional finger model.

The third step is to normalize the three-dimensional finger model to achieve the effect of eliminating the horizontal and vertical offset of the finger. It means that the least square method is used to return the center of the approximate ellipse on each cross-section obtained in the three-dimensional reconstruction to a line. On the central axis, then use the following equation (1) to normalize the coordinates;

Among them, (x _m , y _m , z _m ) represents the midpoint of the ellipse, and (S, W, G) represents the direction of the central axis.

Through the above normalization, the axial direction of the finger can be consistent with the central axis of the three-dimensional model, and the center point of the three-dimensional model can be consistent with the origin, thereby eliminating the offset caused by the horizontal and vertical movements.

In the fourth step, processing the normalized three-dimensional finger model to generate a three-dimensional texture expansion map and geometric distance feature map means:

First, define the sector-shaped cylinder area as SC-Block (i), where i is a subscript from 1 to N; rotate and cut along the axis of the three-dimensional cylinder to obtain 360 sector-shaped cylinder areas; The range of the center angle of the bottom surface of the volume area is set to ((i-1) · Δα, i · Δα]; at the same time, the range of the cylinder height Z is set to [z _min , z _max ], where z _min and z _max represent respectively The minimum and maximum heights; N represents the width of the feature map, N = 360 / Δα, and Δα is the angle sampling interval;

Then, the three-dimensional point set of the fan-shaped cylinder region set is mapped to the three-dimensional texture expansion map I _F3DTM and the geometric distance feature map I _F3DGM by the following function:

I _F3DTM .col (i) = Γ _t (SC-Block (i)) (1)

I _F3DGM .col (i) = Γ _g (SC-Block (i)) (2)

Among them, F3DTM and F3DGM respectively represent the three-dimensional texture expansion map and geometric distance feature map, .col (i) represents the i-th column of the feature map, and the functions Γ _g and Γ _t respectively collect the fan-shaped cylinder area set SC-Block (i ) Divide and cut into M blocks from the Z axis at fixed intervals; wherein, each pixel of _{IF3DTM is} obtained by calculating the average pixel value in the area, and each pixel of _IF3DGM is calculated by calculating the point set in the corresponding area to the central axis The average value of the linear distance is obtained.

In the fourth step, the convolutional neural network is used to extract features of the three-dimensional texture expansion map and the geometric distance feature map to obtain the vein texture feature and the central axis geometric distance feature; at the same time, training the neural network means that the structure of the neural network is It is composed of four convolutional blocks containing 3 × 3 and 1 × 1 convolutional layers continuously stacked, so that the design can effectively reduce the number of parameters while ensuring its recognition performance; the three-dimensional texture expansion map and geometric distance feature map are passed in turn The neural network structure and the 256-dimensional output of the fully connected layer obtain the 256-dimensional vein texture feature and the 256-dimensional central axis geometric distance feature. Finally, the SoftMax layer calculates the loss and trains the network.

The following describes the proposed three-dimensional finger vein matching method:

By calculating the vein texture feature and median geometric distance feature score of the template sample and the sample to be matched, and performing weighted fusion, the fusion matching score is judged by the threshold, and the matching recognition of the three-dimensional finger vein is completed.

Specifically, in the feature matching stage, firstly, the template sample finger vein to be matched and the sample finger vein to be matched are sequentially performed from the first step to the fourth step, respectively, to obtain the vein texture and three-dimensional finger shape features (middle-axis geometry) Distance feature), and the vein texture and three-dimensional finger shape feature of the sample to be matched; the cosine distance D ₁ of the template sample and the vein texture feature of the sample to be matched, and the cosine distance D of the three-dimensional finger shape feature of the template sample and the sample to be matched, respectively ₂ . Their cosine distance formulas are as follows:

F _v1 and F _v2 are the template sample and the vein feature vector of the finger with the matching sample, F _d1 , F _d2 or finger shape feature vector, respectively.

Then, the cosine distance (matching score) of the vein texture feature and the finger shape feature is scored and weighted to obtain the total cosine distance D. Among them, the fusion weight is randomly selected 10% in the data as the verification set, and the weight value is traversed on the verification set, and the weight value with the lowest error rate after the fusion matching score is taken as the best weight, and the best weight is used to match The results are weighted and fused to get the final matching result;

S = w · S _t + (1-w) · S _g

Where S is the final matching score, S _t is the texture matching score, S _g is the shape matching score, and w is the fusion weight.

Finally, a threshold is determined through experiments. When the total cosine distance D is less than the threshold, it is determined as a match, otherwise it does not match.

Compared with the prior art, the present invention has the following advantages and beneficial effects: The three-dimensional finger vein feature extraction method and matching method of the present invention can obtain more vein texture features, get a better matching recognition effect, and can also be effectively solved The change of finger pose brings the problem of poor matching recognition performance, thereby improving the accuracy and effectiveness of vein matching recognition.

BRIEF DESCRIPTION

1 is a schematic flowchart of a three-dimensional finger vein feature extraction method and matching method of the present invention;

2 is a schematic diagram of the construction of the three-dimensional finger model of the ellipse model of the present invention;

3 is a schematic view of 360 sector-shaped cylinder regions obtained by rotating and cutting along the axis of a three-dimensional cylinder of the present invention;

4 is a three-dimensional texture development diagram of the present invention;

Figure 5 is a geometric distance characteristic diagram of the present invention;

6 is a schematic diagram of a fully connected layer through a convolutional neural network structure and a 256-dimensional output of the three-dimensional texture expansion map and geometric distance feature map of the present invention.

detailed description

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

Examples

As shown in FIGS. 1 to 6, a three-dimensional finger vein feature extraction method of the present invention includes the following steps:

Among them, in the second step, by calculating the parameters of the three cameras, the two-dimensional finger vein image is mapped to the three-dimensional model, and the construction of the three-dimensional finger model refers to: the finger profile is approximately regarded as an ellipse, and the three-dimensional finger is equidistant Split into several sections, calculate the contour of each section, and use multiple ellipses with different radii and different positions to approximate the model of the finger; then connect all the contours in series according to the direction of the central axis of the finger to obtain an approximate three-dimensional finger model .

The method for calculating the profile of each section is:

1) Establish the xOy coordinate system (2D-CS) according to the projection centers C ₁ , C ₂ , and C _{3 of the} three cameras, as shown in Figure 2.

2) Determine the ellipse and straight line equations.

The equation for the ellipse is as follows:

The projection center of each camera is recorded as C _i (x, y). In this way, the equations of the straight lines C _i U _i (L _ui ) and C _i B _i (L _ni ) can be obtained. Here we only discuss the existence of straight line slope Case:

L _ui : y = k _ui x + b _ui

L _bi : y = k _bi x + b _bi

Where i = 1, 2, 3, k _ui and b _ui respectively represent the slope and intercept of the straight line L _ui , and k _bi and b _bi respectively represent the slope and intercept of the straight line L _bi .

3) Determine the constraints

To make these parallel lines constraining the straight lines, as shown in Figure 2, make them tangent to the ellipse, assuming the equations for these parallel lines are:

L _ui : y = k _ui x + b _ui + ξ _ui

L _bi : y = k _bi x + b _bi + ξ _bi

From the condition that _Lui , _Lbi and the ellipse are tangent, the following constraint equation can be obtained:

among them:

A _ui = a + bk _ui ² + 2ck _ui

B _ui = (2k _ui b + 2c) (b _ui + ξ _ui ) + ek _ui + d

B _bi = (2k _bi b + 2c) (b _bi + ξ _bi ) + ek _bi + d

C _ui = b (b _ui + ξ _ui ) ² + e (b _ui + ξ _ui ) + f

C _bi = b (b _bi + ξ _bi ) ² + e (b _bi + ξ _bi ) + f

4) Objective optimization function

As mentioned earlier, the ellipse must be very close to the constraint equation, so our goal is to minimize the sum of the distances from the ellipse to all straight lines:

5) Solving algorithm

①Single section ellipse solution

We need to convert the coordinates of the edge points on the image to 2D-CS, the conversion relationship is the following formula:

Where θ _i represents

The angle to the positive direction of the x-axis, i = 1, 2, 3, y _m represents the y value of the optical center of the camera, which is related to the parameters within the camera.

For each ellipse, the objective optimization function is solved by the gradient descent method under the constraints shown in 3). The main problem is how to set the initial iteration point, because a proper initial iteration point plays an extremely important role in speeding up the speed of optimization and finding the global optimal solution. Through a large number of experiments, the method of setting the initial iteration point is determined as follows:

There are five independent variables in an ellipse, including the horizontal and vertical coordinates of the center of the ellipse, the length of the major axis of the ellipse, the eccentricity and the angle of rotation. Our problem can be transformed into calculating the approximate inscribed ellipse of the hexagon. According to Brianchon's theory, A hexagon ABCDEF has an inscribed ellipse if and only if its main diagonal AD, BE and CF intersect at the same point:

[(A × D), (B × E), (C × F)] = 0

In the calculation process of our model, these diagonal lines generally intersect each other. We set the initial center point (C ₀ ) of the ellipse as the key point of the triangle formed by these intersection points. The calculation is as shown in the formula below. Set the initial length The axis length is the minimum distance from the initial center point to the six vertices of the hexagon.

In addition, we set the initial eccentricity and rotation angle to fixed values, where the eccentricity is set to e ₀ = 1.4, and the rotation angle is set to α = 0. These two constant values are determined by experiment.

②Reconstruction of the entire 3D finger model

Considering that the more sections are calculated, the more accurate three-dimensional finger model can be obtained, but if the ellipse approximation method is used to calculate more ellipses, it will consume more time, and in practical applications, it is not desirable to reconstruct the three-dimensional finger It took too much time to observe. Through observation, we found such a detail that the change in the axial direction of the finger surface is relatively smooth. We can first select some sparse sets of edge points in the axial direction to reconstruct part of the ellipse, and then use the interpolated The method expands the number of approximate ellipses. However, there is another problem. If the selected edge point set has poor image quality or noise at some edges, resulting in a relatively large error in the position of the detected edge, then the reconstructed finger model There will be relatively large defects in some places. In order to reduce the impact of this aspect and weigh the reconstruction accuracy and time loss, we propose a more robust algorithm to build a three-dimensional model of the finger:

Suppose that in each corrected image, the area with the abscissa in the range of 91 to 490 is set as the effective area. According to the direction of the horizontal axis, the effective area is equally divided into N sub-areas. In each sub-area, we select a set of edge points to reconstruct the ellipse. After obtaining the ellipses of all sub-areas, we use interpolation algorithm to expand the ellipse data. Finally, the ellipse in the two-dimensional plane is converted to the three-dimensional space.

We set the z-coordinate value of an ellipse to the same value without changing the equation of the two-dimensional ellipse. K is the corrected camera parameter.

Among them, (x _m , y _m , z _m ) represents the midpoint of the ellipse, and (S, W, G) represents the direction of the central axis. Through the above normalization, the axial direction of the finger can be consistent with the central axis of the three-dimensional model, and the center point of the three-dimensional model can be consistent with the origin, thereby eliminating the offset caused by the horizontal and vertical movements.

First, define the sector-shaped cylinder area as SC-Block (i), where i is a subscript from 1 to N; rotate and cut along the axis of the three-dimensional cylinder to obtain 360 sector-shaped cylinder areas, as shown in Figure 3 As shown. Set the range of the center angle of the bottom of the fan-shaped cylinder area to ((i-1) · Δα, i · Δα]; meanwhile, set the range of the cylinder height Z to [z _min , z _max ], where z _min and z _max represents the minimum and maximum values of height respectively; N represents the width of the feature map, N = 360 / Δα, and Δα is the angle sampling interval;

I _F3DTM .col (i) = Γ _t (SC-Block (i)) (1)

I _F3DGM .col (i) = Γ _g (SC-Block (i)) (2)

Among them, F3DTM and F3DGM respectively represent the three-dimensional texture expansion map and geometric distance feature map, .col (i) represents the i-th column of the feature map, and the functions Γ _g and Γ _t respectively collect the fan-shaped cylinder area set SC-Block (i ) Divide and cut into M blocks from the Z axis at fixed intervals; wherein, each pixel of _{IF3DTM is} obtained by calculating the average pixel value in the area, and each pixel of _IF3DGM is calculated by calculating the point set in the corresponding area to the central axis The average value of the linear distance is obtained. In this embodiment, Δα = 1 and M = 360. Figures 4 and 5 are examples of calculated feature maps.

In the fourth step, the convolutional neural network is used to extract features of the three-dimensional texture expansion map and the geometric distance feature map to obtain vein texture features and central axis geometric distance features; meanwhile, training the neural network means: as shown in Figure 6 , The neural network structure is composed of four convolutional blocks containing 3 × 3 and 1 × 1 convolutional layers stacked in succession, so that the design can effectively reduce the number of parameters while ensuring its recognition performance; 3D texture expansion map and geometry The distance feature map respectively passes through the neural network structure and the 256-dimensional output fully connected layer to obtain the vein texture feature and the central axis geometric distance feature; finally, the SoftMax layer calculates the loss and trains the network.

The three-dimensional finger vein feature matching method of the present invention is as follows: by calculating the vein texture feature and mid-axis geometric distance feature score of the template sample and the sample to be matched, and performing weighted fusion, the fusion score is determined by a threshold to complete Three-dimensional finger vein matching recognition.

Specifically, in the feature matching stage, the first step and the fourth step are performed on the finger veins of the template sample to be matched and the finger veins of the sample to be matched, respectively, to obtain the vein texture and three-dimensional finger shape features of the template sample, and to be matched The vein texture and three-dimensional finger shape features of the sample; the cosine distance D _{1 of the} vein texture feature of the template sample and the sample to be matched, and the cosine distance D ₂ of the three-dimensional finger shape feature of the template sample and the sample to be matched are calculated respectively. Their cosine distance formulas are as follows:

Among them, F _v1 and F _v2 are the template sample and the vein feature vector of the finger with the matching sample, and F _d1 and F _d2 are the shape feature vector of the template sample and the finger with the matching sample, respectively.

After that, the cosine distance (matching score) of the vein texture feature and the finger shape feature is scored and weighted to obtain the total cosine distance D. Among them, the fusion weight is randomly selected 10% in the data as the verification set, and the weight value is traversed on the verification set, and the weight value with the lowest error rate after the fusion matching score is taken as the best weight, and the best weight is used to match The results are weighted and fused to get the final matching result;

S = w · S _t + (1-w) · S _g

The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, changes, modifications, substitutions, combinations, etc. made without departing from the spirit and principle of the present invention The simplifications should all be equivalent replacement methods, which are all included in the protection scope of the present invention.

Claims

A three-dimensional finger vein feature extraction method, characterized in that it includes the following steps:

In the first step, three fingers are used to capture the finger vein from three angles at equal angles to obtain a two-dimensional finger vein image;

In the second step, by calculating the parameters of the three cameras, the two-dimensional finger vein image is mapped to the three-dimensional model to realize the construction of the three-dimensional finger model;

The third step is to normalize the three-dimensional finger model to achieve the effect of eliminating the horizontal and vertical offset of the finger;

In the fourth step, feature extraction is performed on the normalized 3D finger model:

(1) Process the normalized three-dimensional finger model to generate three-dimensional texture expansion map and geometric distance feature map;

(2) The convolutional neural network is used to extract features from the three-dimensional texture expansion map and the geometric distance feature map to obtain vein texture features and the central axis geometric distance features; at the same time, the neural network is trained.
The three-dimensional finger vein feature extraction method according to claim 1, characterized in that in the second step, by calculating the parameters of the three cameras, the two-dimensional finger vein image is mapped to the three-dimensional model to realize the construction of the three-dimensional finger model means : The finger profile is approximately regarded as an ellipse, the three-dimensional finger is divided into several sections at equal distances, the contour of each section is calculated, and multiple ellipses with different radii and different positions are used to approximate the model of the finger; The three-dimensional finger model can be obtained by connecting the fingers in the axial direction in series.
The three-dimensional finger vein feature extraction method according to claim 1, characterized in that: in the third step, the three-dimensional finger model is normalized to eliminate the influence of the horizontal and vertical displacement of the finger means: using the least square method Return the center of the approximate ellipse on each cross-section obtained in the three-dimensional reconstruction to a central axis, and then use the following equation (1) to normalize the coordinates;

Among them, (x m , y m , z m ) represents the midpoint of the ellipse, and (S, W, G) represents the direction of the central axis.
The three-dimensional finger vein feature extraction method according to claim 1, wherein in the fourth step, the normalized three-dimensional finger model is processed to generate a three-dimensional texture expansion map and a geometric distance feature map means:

First, define the sector-shaped cylinder area as SC-Block (i), where i is a subscript from 1 to N; rotate and cut along the axis of the three-dimensional cylinder to obtain 360 sector-shaped cylinder areas; The range of the center angle of the bottom of the volume area is set to ((i-1) · △ α, i · △ α]; meanwhile, the range of the cylinder height Z is set to [z min , z max ], where z min and z max Respectively represent the minimum and maximum values of the height; N represents the width of the feature map, N = 360 / Δα, Δα is the angle sampling interval;

Then, the three-dimensional point set of the fan-shaped cylinder region set is mapped to the three-dimensional texture expansion map I F3DTM and the geometric distance feature map I F3DGM by the following function:

I F3DTM .col (i) = Γ t (SC-Block (i)) (1)

I F3DGM .col (i) = Γ g (SC-Block (i)) (2)

Among them, F3DTM and F3DGM respectively represent the three-dimensional texture expansion map and geometric distance feature map, .col (i) represents the i-th column of the feature map, and the functions Γ g and Γ t respectively collect the fan-shaped cylinder area set SC-Block (i ) Divide and cut into M blocks from the Z axis at fixed intervals; wherein, each pixel of IF3DTM is obtained by calculating the average pixel value in the area, and each pixel of IF3DGM is calculated by calculating the point set in the corresponding area to the central axis The average value of the linear distance is obtained.
The three-dimensional finger vein feature extraction method according to claim 1, characterized in that: in the fourth step, a convolutional neural network is used to perform feature extraction on the three-dimensional texture expansion map and the geometric distance feature map, respectively, to obtain the vein texture feature and the central axis Geometric distance feature; training the neural network at the same time means that the neural network structure is formed by continuously stacking four convolution blocks containing 3 × 3 and 1 × 1 convolutional layers; three-dimensional texture expansion map and geometric distance feature map Through the neural network structure and the 256-dimensional output fully connected layer in turn, the 256-dimensional vein texture feature and the 256-dimensional geometric axis geometric distance feature are obtained; finally, the SoftMax layer calculates the loss and trains the network.
The three-dimensional finger vein feature matching method according to claim 1, characterized in that: by calculating the vein texture feature and the median geometric distance feature score of the template sample and the sample to be matched, and performing weighted fusion, the fusion The matching score is judged to complete the matching recognition of the three-dimensional finger veins.
The three-dimensional finger vein feature matching method according to claim 6, characterized in that: the vein texture feature and the median geometric distance feature score of the template sample and the sample to be matched are calculated, and weighted fusion is performed, and the fusion is performed by a threshold After the matching score is judged, the completion of the three-dimensional finger vein matching recognition is:

In the feature matching stage, the first step and the fourth step are performed on the finger vein of the template sample to be matched and the finger vein of the sample to be matched, respectively, to obtain the vein texture and three-dimensional finger shape characteristics of the template sample, and the vein of the sample to be matched Texture and three-dimensional finger shape features; calculate the cosine distance D 1 of the template sample and the vein texture features of the sample to be matched, and the cosine distance D 2 of the three-dimensional finger shape features of the template sample and the sample to be matched, and the formulas of the cosine distance are as follows:

Among them, F v1 and F v2 are the template sample and the vein feature vector of the finger with the matching sample, and F d1 and F d2 are the shape feature vector of the template sample and the finger with the matching sample, respectively.
The three-dimensional finger vein feature matching method according to claim 7, wherein: the cosine distance between the vein texture feature and the finger shape feature is fractionally layered and weighted to obtain the total cosine distance D; wherein, the fusion weight is obtained by randomizing the data Extract 10% as the verification set, traverse the weight values on the verification set, take the weight value with the lowest error rate after merging the matching scores as the optimal weight, and use this optimal weight to fuse the matching results to obtain the final matching result ;

S = w · S t + (1-w) · S g

Where S is the final matching score, S t is the texture matching score, S g is the shape matching score, and w is the fusion weight;

Finally, a threshold is determined through experiments. When the total cosine distance D is less than the threshold, it is determined as a match, otherwise it does not match.