A high-security identity-oriented revocable finger vein detector

Abstract

The application provides a high-security identity recognition-oriented revocable finger vein detector, comprising a BWR-ROIAlign plug-and-play module which is composed of three sub-modules of positioning, compression and conversion. The CFVD is divided into two stages of shallow feature extraction and deep feature extraction, and specifically divided into five steps, including: step 1, performing size normalization on the original finger vein image and extracting shallow features; step 2, acquiring ROI coordinate information by using the positioning sub-module; step 3, removing spatial redundancy in the shallow features by using the compression sub-module; step 4, removing channel redundancy by using De-R Conv to compress the channels of the shallow features after removing the spatial redundancy; and step 5, encrypting the feature map by block warping remapping of the conversion sub-module, then extracting protected deep features, and finally completing high-security identity recognition in the encryption domain. The scheme is efficient, safe and intelligent, can realize revocation and reuse of finger vein biometric features, and can achieve an average recognition accuracy of 99.5%.

Description

Technical Field

[0001] This invention relates to a revocable finger vein detector for high-security identity recognition, belonging to the field of biometric technology. Background Technology

[0002] With the development of the information society, traditional authentication technologies based on passwords or ID cards can no longer meet people's security needs. In recent years, biometrics has shown many advantages in applications in finance, security, and healthcare. Biometrics refers to the automatic capture, processing, analysis, and recognition of digital physiological or behavioral characteristic signals of the human body by intelligent machines. This is a typical pattern recognition task and has always been at the forefront of artificial intelligence development, mainly including face recognition, fingerprint recognition, vein recognition, voice recognition, and gait recognition. Finger vein recognition, as a representative technology of second-generation biometrics, has become the mainstream of recent research due to its convenience and higher security. Finger vein recognition is mainly divided into image acquisition, preprocessing, feature extraction, and recognition authentication. Among them, deep learning-based finger vein recognition systems will automatically complete the extraction and recognition of vein features in a unified manner. Preprocessing, as a key step in finger vein recognition, can effectively improve system performance. However, due to the different structures of acquisition equipment and application environments, special image processing operations (region of interest extraction, image enhancement, and feature alignment, etc.) need to be designed accordingly. This separates the preprocessing operation from the subsequent feature extraction and recognition authentication, seriously affecting the convenience, integrity, and automation level of the recognition system.

[0003] Meanwhile, with the widespread application of biometric technology, the information leakage problems it brings seriously damage users' privacy and anonymity. Although biometric features are more difficult to fraudulently copy or forge than traditional authentication technologies, and improve convenience and security, the unique nature of biometric features allows attackers to track the activities of subjects registered in different domains. Once a biometric feature is leaked, the security of the identification system drops sharply, and the biometric feature becomes unusable, further limiting the biometric features a subject can use. These issues demonstrate that biometric information cannot be revoked or republished. To address this problem, the concept of revocable biometrics has been introduced. This involves intentionally and repeatedly distorting the biometric signal, allowing the biometric template to be authenticated in an encrypted domain, thus protecting the biometric feature.

[0004] A good revocable biometric identification technology should possess the following characteristics: (1) Irreversibility: Given a protected template, it should be impossible to reconstruct the original biometric sample; (2) Reversibility: Multiple protected templates can be generated from a given biometric sample; (3) Non-linkability: If two protected templates are generated from the same biometric information and stored in different systems, it is impossible to determine that they belong to the same object; (4) Performance: Using the BTP scheme should not significantly reduce the system's recognition performance. Furthermore, the recognition performance should not be sensitive to the parameters of the specified template protection steps. Unlike traditional identification and authentication methods, finger vein recognition systems based on Deep Convolutional Neural Networks (DCNNs) do not require storing feature templates in a database, thus avoiding the problem of template theft. However, attackers can intrude into the finger vein system to analyze the finger vein information output by the network. The main component of DCNN is convolution operation, and the original vein template can be recovered to a certain extent through deconvolution. This leads to a security risk for DCNN-based finger vein recognition systems caused by template recovery. To address these issues, this invention combines biometric template protection with a DCNN-based finger vein recognition system, designing a template-protected recognition method. It utilizes simple and effective irreversible transformations to reconstruct or distort biometric features, and employs revocable biometric template protection, preventing attackers from reconstructing the original biometrics using the encrypted template. To solve these problems and achieve intelligent, integrated, secure finger vein recognition, this invention provides a revocable finger vein detector for constructing a recognition system, designed for high-security identity verification. Summary of the Invention

[0005] Currently, many revocable finger vein recognition systems separate the template protection process from the recognition process. Typically, the original feature template is first converted to a protected domain and stored in a database, to be transmitted later or fed into a classifier for recognition. Furthermore, the original feature template often requires sophisticated preprocessing, including region of interest extraction and feature alignment. This makes revocable finger vein recognition systems cumbersome and highly unreliable.

[0006] To address the aforementioned technical problems, this invention employs the following technical means: a fully automatic revocable finger vein detector (CFVD). It primarily comprises a BWR-ROIAlign plug-and-play module, which consists of three sub-modules: localization, compression, and transformation. CFVD uses ResNet50, commonly used in classification tasks, as its backbone. The entire CFVD model is divided into two stages based on the vein features processed in different domains: a shallow feature (source domain) extraction stage and a deep feature (encrypted domain) extraction stage.

[0007] The specific steps are as follows:

[0008] Step 1: The acquired raw finger vein image is normalized in size (usually to 256×320) and denoted as F. F is then input into the CFVD shallow feature extraction stage to obtain the finger vein shallow feature F. S .

[0009] Step 2: Transfer the shallow features F S Inputting the localization submodule of BWR-ROIAlign, the four regression coordinates of the ROI are obtained using shallow features, resulting in the ROI coordinate information P(x min ,y min ,x max ,y max ).

[0010] Step 3: Transfer the shallow features F S The compression submodule in BWR-ROIAlign first uses the ROI coordinate information P to extract the superficial features F of the finger veins. S A unique stable region F is delineated in the middle. ROI To achieve spatial redundancy compression, the next step will be to obtain F. ROI Channel redundancy removal is performed using the De-R Conv method provided in this invention to obtain a compact finger vein feature map F. c .

[0011] Step 4: Process the feature map F obtained in the shallow feature processing stage. c Input deep feature (encrypted domain) extraction and analysis stage. First, F c In the transformation submodule of BWR-ROIAlign, the proposed Block Warping Remapping (BWR) template protection method is used to transform F... c Convert to F p .

[0012] Step 5: Convert the feature map F p The finger vein feature information is extracted and analyzed after deep feature extraction and encryption. Finally, the extracted feature information is input into a fully connected layer for recognition and classification.

[0013] Furthermore, in step 2, for finger ROI localization, ROI coordinate regression involves finding a mapping f such that... Thus reduce Compared with the true bounding box (G x G y G w G h The offset between () for a total of N training pairs. A set of.

[0014]

[0015] Where t is the offset between the regression label and the suggestion box. * represents four types of mapping: translation transformations Δx and Δy, scaling transformation S. w and S h F i This is the input feature map. Finally, ω is learned by optimizing the regularized least squares objective, where λ is the regularization coefficient.

[0016]

[0017] However, the ROI in finger vein images is not a small objective. The localization submodule predicts the offset between bounding boxes by learning four cumbersome mappings, while directly predicting the true coordinates. The objective function is defined as:

[0018]

[0019] Where C represents the true coordinates (x, y) of the ROI in the finger vein image. min ,y min ,x max ,y max Similarly, F i It is the input feature map. Predicted ROI coordinates obtained from feature maps ω is learned by optimizing the regularized least squares objective.

[0020] Furthermore, in step 3, the coordinates P(x) of the ROI rectangular window provided in the positioning submodule are... min ,y min ,x max ,y max The mapped features are then mapped onto the corresponding grid cells on the feature map. These mapped grid cells are then divided into H×W sub-regions. Finally, a uniform-sized ROI feature map F is sampled from these sub-regions. ROI The dimensions are (B, C, H, W). It is important to note that linear interpolation is used in all possible quantization operations to avoid any quantization loss in coordinate mapping and mesh generation. Further, in step 3, De-R Conv obtains the feature F... ROI It is divided into 4 groups, and each group is merged to reduce redundant feature maps. It is divided into two paths, positive and negative.

[0021] First, consider a feature map F with scale (B, C, H, W). ROI The weights w and reverse weights of each channel are obtained using average pooling and two 1×1 convolutions. This is an operation similar to channel attention.

[0022] Next, the weights w and w′ are assigned to the corresponding feature maps to obtain F. w and F w and They were divided into four groups and linearly fused to obtain scales of [scale value missing]. Feature map F w 'and

[0023]

[0024]

[0025] in, This means dividing the weighted feature map into four groups of non-overlapping feature maps. This indicates that the corresponding positions of the four sets of feature maps will be linearly fused.

[0026] Next, F w 'and Each feature map is fed into a standard 3×3 convolution to extract dense feature maps.

[0027] Finally, the feature map size obtained from the forward path portion is... The feature map obtained from the reverse path portion has a size of [size missing]. The feature maps obtained from the forward and reverse paths are concatenated to integrate the complete information, resulting in a size of [size missing]. Feature map F c As output.

[0028] Furthermore, in step 4, a method for protecting biometric templates by block distortion remapping is provided, which is mainly divided into two parts: block grid distortion and block remapping, and has four predefined hyperparameters: image block size b, distortion grid size s, distortion factor o, and resampling rate r, to balance the security and recognition performance of the recognition system.

[0029] For a source region feature map F c (x, y), with dimensions (H, W). Let F... c Divide into pixel blocks of size b×b, denoted as A. T T = 1, 2, ..., (HW / b) 2 ), continue with A T Divide the grid into quadrilaterals of size s×s in the X and Y directions, and then apply the grid to A. T Randomly transform all the vertices of the grid to obtain the distorted feature patch B. T For any given grid, Axi A yi Let A be the i-th vertex. i The x and y coordinates of B. i For the i-th vertex after deformation, B is similarly... xi B yi Let be the x and y coordinate values ​​after the transformation. Then, the coordinate relationship between the vertices before and after the transformation is:

[0030]

[0031] Where Δx and Δy are the lengths of a single grid cell in the X and Y directions, respectively. x o y These are the distortion factors in the X and Y directions, respectively, with values ​​ranging from [0,1]. The closer to 1, the higher the degree of distortion, and vice versa.

[0032] After the vertices are transformed, all A T The area is divided into irregularly shaped grids by random deformation. The next step is to map these grids onto new feature maps C. T In the corresponding rule-based mesh. Assuming a distorted mesh is B1B2B3B4, it has four edges: B1B2, B2B3, B3B4, and B4B1. Dividing these four edges into t-1 equal parts, then B... T The image will be divided into t×t mapping nodes, and the new image will also be divided into t×t nodes. The t-th node in the B1B2 direction... x The t-th node in the direction of node line M1M2, B2B3 y Line N1N2, t x ,t y Given that P is the intersection of M1M2 and N1N2, and ∈ [0,t], then the x and y coordinates of P are... x ,P y for:

[0033]

[0034] M 1x M 1y M1 represents the x and y coordinates; M 2x M 2y Let P be the x and y coordinates of M2. Next, map point P to point P′ at the same location in the new feature map, then the x and y coordinates of P′ are P. x ′,P y 'for:

[0035]

[0036] Where C is the top-left vertex of the new feature patch in the grid, C x C yLet be the x and y coordinates of C. Δt represents the length between two points after the new feature map is equally divided. The twisted mesh points (P) x ,P y The pixel values ​​in (P′) are mapped to the new grid. x ,P′ y The distorted feature patch can be obtained at point (). For C T The set is sampled at a sampling rate of r, and C is selected. T A subset C T Reconstruct the image, and use C T Randomly mapped to feature map F p The final encrypted domain feature map is generated (note that C...). T Some blocks in the ' will be reused, specifically as follows:

[0037] F p =f map (f r (C T ,r),K)

[0038] f r (·) indicates random sampling of the distorted image patch. K is a user-defined external key, f map (·) indicates that f r The subset C obtained after the (·) operation T Using K random sampling, and then mapping to generate an encrypted feature map F. p )

[0039] Compared with existing technologies, the beneficial results of this invention are as follows: Unlike previous finger vein recognition systems, this invention integrates preprocessing and template protection tasks into a deep learning model, proposing a highly efficient, secure, and intelligent integrated finger vein recognition scheme. It is the first end-to-end revocable finger vein recognition system. This invention can automatically locate the region of interest (ROI) of finger veins, obtain stable and unique vein regions, reduce manpower consumption, lower costs, and improve recognition efficiency. Furthermore, the vein feature information during the recognition process is highly irreversible, unlinkable, and revocable, effectively solving the privacy leakage and security problems inherent in finger vein recognition systems. Simultaneously, the BWR-ROIAlign provided by this invention is a plug-and-play architectural unit that can introduce the above characteristics into DCNN-based finger vein recognition systems. This invention provides a novel method and approach for a highly efficient, secure, and intelligent integrated design scheme for finger vein recognition, enabling the revocable reuse of finger vein biometric features and achieving an average recognition accuracy of 99.5%. In terms of security, CFVD-based finger vein recognition can effectively avoid brute-force attacks, reversal attacks, cross-cutting attacks, and dictionary attacks, and possesses excellent revocable biometric features. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the implementation examples will be briefly introduced below. Obviously, the accompanying drawings shown below are only individual examples of the present invention. For researchers in the art, this method can be applied to finger vein images of different qualities by simply reproducing the present invention.

[0041] Figure 1 This is an overall structural diagram of the CFVD described in this invention.

[0042] Figure 2 This is a schematic diagram of the positioning submodule in the BWR-ROIAlign unit described in this invention.

[0043] Figure 3 This is a schematic diagram of feature alignment to avoid quantization loss as described in this invention.

[0044] Figure 4 This is a schematic diagram of the De-R Conv method described in this invention.

[0045] Figure 5 This is a schematic diagram of the conversion submodule in the BWR-ROIAlign unit described in this invention.

[0046] Figure 6 This is a rendering of the positioning submodule provided by the present invention.

[0047] Figure 7 This is a schematic diagram showing the comparison of recognition performance obtained from the simulation test of this invention.

[0048] Figure 8 This is a comparative diagram of the safety performance obtained from the simulation test of this invention. Detailed Implementation

[0049] To make the above-mentioned objectives, features, and advantages of the present invention more apparent and understandable, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that these embodiments are only a part of the examples of the present invention, not all of them. All other examples obtained without innovative work are within the protection scope of the present invention. Example 1: A fully automatic revocable finger vein detector (CFVD), its overall structure is as follows... Figure 1 As shown. CFVD uses ResNet50, commonly used in classification tasks, as its backbone. The entire CFVD model is divided into two stages based on processing vein features from different domains: a shallow feature (source domain) extraction stage and a deep feature (encrypted domain) extraction stage; the specific steps are as follows:

[0050] In the source domain feature extraction stage, the obtained superficial features of the finger veins often contain location, shape, or noise information, but lack deep semantic information for discrimination. At this stage, different samples of the same instance exhibit significant inter-class differences. However, these superficial features can be used for location regression of the region of interest (ROI) of the finger veins. Therefore, the localization submodule in the BWR-ROIAlign module utilizes these superficial features to obtain the four regression parameters needed for ROI extraction. The acquired original finger vein image is normalized, and its resulting F is input into the CFVD superficial feature extraction stage to obtain the superficial finger vein feature F. S .

[0051] First, the shallow features F S Inputting the localization submodule of BWR-ROIAlign, the four regression coordinates of the ROI are obtained using shallow features, resulting in the ROI coordinate information P(x min ,y min ,x max ,y max ).like Figure 2 As shown.

[0052] Next, the shallow features F S The compression submodule in BWR-ROIAlign first uses the ROI coordinate information P to extract the superficial features F of the finger veins. S A unique stable region F is delineated in the middle. ROI This achieves spatial redundancy compression. Specifically, it involves adjusting the coordinates P(x) of the ROI rectangular window provided in the positioning submodule. min ,y min ,x max ,y max The mapped features are then mapped onto the corresponding grid cells on the feature map. These mapped grid cells are then divided into H×W sub-regions. Finally, a uniform-sized ROI feature map F is sampled from these sub-regions. ROI The dimensions are (B, C, H, W). It is important to note that linear interpolation should be used in all operations where quantization may occur to avoid any quantization loss in coordinate mapping and mesh generation. For example... Figure 3 As shown.

[0053] Next, channel redundancy is removed, and De-R Conv obtains the feature F. ROI Divided into 4 groups, each group is merged to reduce redundant feature maps. It consists of two paths, forward and reverse, such as... Figure 4 As shown. First, for a feature map F with scale (B, C, H, W) ROI The weights w and reverse weights of each channel are obtained using average pooling and two 1×1 convolutions. This is an operation similar to channel attention.

[0054] Then, the weights w and w′ are assigned to the corresponding feature maps to obtain F. w and F w and They were divided into four groups and linearly fused to obtain scales of [scale value missing]. Feature map F w 'and use in, This means dividing the weighted feature map into four groups of non-overlapping feature maps. This indicates that the corresponding positions of the four feature maps will be linearly fused. Then F... w 'and Each feature map is fed into a standard 3×3 convolution to extract dense feature maps.

[0055] Finally, the feature map size obtained from the forward path portion is... The feature map obtained from the reverse path portion has a size of [size missing]. The feature maps obtained from the forward and reverse paths are concatenated to integrate the complete information, resulting in a size of [size missing]. Feature map F c As output.

[0056] The next stage is deep feature extraction, which mainly utilizes the BWR-ROIAlign module to convert shallow features into encrypted deep features. For a source domain feature map F... c (x, y), with dimensions (H, W). Let F... c Divide into pixel blocks of size b×b, denoted as A. T T = 1, 2, ..., (HW / b) 2 ), continue with A T Divide the grid into quadrilaterals of size s×s in the X and Y directions, and then apply the grid to A. T Randomly transform all the vertices of the grid to obtain the distorted feature patch B. T After the vertices are transformed, all A's T The area is divided into irregularly shaped grids by random deformation. The next step is to map these grids onto new feature maps C. T In the corresponding rule's grid. Next, map point P to the point P at the same location in the new feature map. ′ Then P ′ The x and y coordinates P x ′,P y 'for: Where C is the top-left vertex of the new feature patch in the grid, C x C yLet be the x and y coordinates of C. Δt represents the length between two points after the new feature map is equally divided. The twisted mesh points (P) x ,P y The pixel values ​​in (P′) are mapped to the new grid. x ,P′ y The distorted feature patch can be obtained at point (). For C T The set is sampled at a sampling rate of r, and C is selected. T A subset C T Reconstruct the image, and use C T Randomly mapped to feature map F p The final encrypted domain feature map is generated (note that C...). T (Some blocks in ' will be reused), F p =f map (f r (C T ,r),K), where f r (·) indicates random sampling of the distorted image patch. K is a user-defined external key, f map (·) indicates that f r The subset C obtained after the (·) operation T Using K random sampling, and then mapping to generate an encrypted feature map F. p ).

[0057] Finally, the transformed feature map F p The finger vein feature information is extracted and analyzed after deep feature extraction and encryption. Finally, the extracted feature information is input into a fully connected layer for recognition and classification.

[0058] The above-described embodiments merely illustrate some implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A revocable finger vein detector for high-security identity recognition, characterized in that, The detector includes a BWR-ROIAlign plug-and-play module, which consists of three sub-modules: localization, compression, and transformation. CFVD uses ResNet50 from the classification task as the skeleton. The entire CFVD model is divided into two stages according to the vein features processed in different domains: the shallow feature extraction stage and the deep feature extraction stage. The solution is as follows: Step 1: Normalize the size of the acquired raw finger vein image and record it as... ,Will The superficial features of finger veins are obtained from the CFVD shallow feature extraction stage. , Step 2: Shallow features Inputting the localization submodule of BWR-ROIAlign, the four regression coordinates of the ROI are obtained using shallow features, resulting in the ROI coordinate information. , Step 3: Add shallow features The compression submodule in BWR-ROIAlign first utilizes the ROI coordinate information P to extract the superficial features of the finger veins. A unique stable region was delineated in the middle. To achieve spatial redundancy compression, the following will be obtained Channel redundancy removal was performed using the De-R Conv method to obtain a compact finger vein feature map. , Step 4: Process the feature map obtained from the shallow feature processing stage. In the deep feature extraction and analysis stage, the first step is to... In the transformation submodule of BWR-ROIAlign, the block warp remapping template protection method is used to... Convert to , Step 5: Convert the feature map The finger vein feature information after deep feature extraction and analysis of the encrypted domain is input, and finally the extracted feature information is input into the fully connected layer for recognition and classification. In step 3, De-R Conv will obtain the features It is divided into 4 groups, and each group is merged to reduce redundant feature maps. It consists of two paths, forward and reverse. First, for a scale of Feature map Using average pooling and two Convolution obtains the weights for each channel. and reverse weights This is an operation similar to channel attention. Next, we will discuss the weights. and Assigning the corresponding feature map yields and ,Will and They were divided into four groups and linearly fused to obtain scales of [scale value missing]. Feature map and , in, This means dividing the weighted feature map into four groups of non-overlapping feature maps. This indicates that the corresponding positions of the four sets of feature maps are linearly fused. Next and Send one standard Convolution extracts features from dense feature maps. Finally, the feature map size obtained from the forward path portion is... The feature map obtained from the reverse path portion has a size of [size missing]. The feature maps obtained from the forward and reverse paths are connected to integrate the complete information, resulting in a size of [size missing]. Feature map As output; Step 4 presents a method for biometric template preservation using block warping and remapping, which mainly consists of two parts: block mesh warping and block remapping, and has four predefined hyperparameters: image block size. Twisted mesh size Distortion factor and resampling rate This is used to balance the security and recognition performance of the identification system. For a source region feature map Its size is ,Will Divided into sizes A pixel block of size, denoted as , Continue exist Mesh division in direction The size of the quadrilateral, then... Randomly transform all the vertices of the grid to obtain distorted feature patches. For any given grid, , For the first vertices The x and y coordinate values, and For the deformed first Each vertex, similarly , Let be the x and y coordinate values ​​after the transformation. Then, the coordinate relationship between the vertices before and after the transformation is: in , Each individual grid in Length in direction , They are respectively The distortion factor in the direction, with a value range of The closer The higher the degree of distortion, the lower the degree of distortion. After the vertices are transformed, all The area is divided into irregularly shaped grids with random deformations; the next step is to map these grids onto new feature maps. In the corresponding rule-based mesh, let a certain distorted mesh be... Then there are a total of Four edges, perform the following steps on the four edges. Divide into equal parts, then Will be divided into Each mapping node also divides the new image into... 1 node, Direction first Strip node line , Direction First Strip node line , , for and The intersection point, then x and y coordinates , for: , for x and y coordinates; , for The x and y coordinates, next, the points Points mapped to the same location in the new feature map ,but x and y coordinates for: （10） in The top-left vertex of this grid is the new feature patch. ,for The x and y coordinates, The length between two points after the new feature map is divided equally will be represented by the distorted grid points. The pixel values ​​in ) are mapped to the new grid. The distorted feature patch can be obtained at point ) and for The set sampling rate Sampling, selection A subset Reconstructing the image, Random mapping to feature map The final encrypted domain feature map is generated as follows: This indicates that random sampling is performed on the distorted image patch. For user-defined external keys, Indicates to The subset obtained after the operation use Random sampling is performed, and then the encrypted feature map is generated. .

2. The revocable finger vein detector for high-security identity recognition according to claim 1, characterized in that, Step 2 is described in detail below: For finger ROI localization, ROI coordinate regression is about finding a mapping. Make Thus reducing With the true bounding box The offset between, for common training pairs The set, in It is the offset between the regression labels and the suggestion boxes. Representing four types of mappings: translation transformation and Scale transformation and , The input feature map is used to learn the final regularized least squares objective. ,in It is the regularization coefficient. However, the ROI in finger vein images is not a small objective. The localization submodule predicts the offset between bounding boxes by learning four cumbersome mappings, while directly predicting the true coordinates. The objective function is defined as: in Represents the true coordinates of the ROI in the finger vein image. Similarly, It is the input feature map. Predicted ROI coordinates obtained from feature maps ,in Learning is achieved by optimizing the regularized least squares objective.

3. The revocable finger vein detector for high-security identity recognition according to claim 2, characterized in that, Step 3 is described in detail below: The coordinates of the ROI rectangular window provided in the positioning submodule Mapping to the corresponding grid cells on the feature map, and then dividing the mapped grid cells into... A number of sub-regions are selected, and finally, a ROI feature map of uniform size is obtained by sampling from these sub-regions. The size is .

4. A terminal, characterized in that: The system includes a processor, a memory, and a revocable finger vein detector algorithm program for high-security identity recognition stored in the memory. When the processor runs the revocable finger vein detector program for high-security identity recognition, it implements the steps of the revocable finger vein detector for high-security identity recognition as described in any one of claims 1-3.

5. A computer-readable storage medium storing computer instructions thereon, characterized in that: When executed by a processor, the computer instructions implement the steps of the revocable finger vein detector for high-security identity recognition as described in any one of claims 1-3.

Images