Fingerprint activity detection method based on group feature analysis

By using the GAFA-Net detection model based on group feature analysis, the problems of low efficiency and low accuracy in fingerprint activity detection systems are solved, and more efficient forgery fingerprint identification and classification is achieved.

CN118262387BActive Publication Date: 2026-11-17CHINA THREE GORGES UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410219254.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-27
Publication Date
2026-11-17
Estimated Expiration
2044-02-27

AI Technical Summary

Technical Problem

In existing fingerprint activity detection systems, network detection efficiency is low, the accuracy of identifying and classifying forged fingerprints is not precise enough, and there is a lack of mining of deep texture features of fingerprint information.

Method used

The GAFA-Net detection model based on group feature analysis, including a multi-attention mechanism, is adopted. The GAFA-Net detection model is built by preprocessing the fingerprint dataset. The group feature analysis module is added to the EfficientNetV2-S model for discrimination and classification.

Benefits of technology

It improves the accuracy of forged fingerprint identification and classification, reduces computational resource consumption, has greater applicability, and can maintain good results even when training data is mismatched.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118262387B_ABST
    Figure CN118262387B_ABST
Patent Text Reader

Abstract

This invention provides a fingerprint activity detection method based on group feature analysis, belonging to the technical field of biometric recognition. The invention includes: acquiring a fingerprint dataset and performing image preprocessing; constructing a GAFA-Net (Group Feature Analysis Fusion Network) detection model; dividing the dataset into a training set, training the model using the training set, and saving the training set; preprocessing the images to be detected, inputting them into the optimal model, and having the optimal model classify the images. This invention is used to distinguish between forged and genuine fingerprints. Based on the EfficientNetV2-S model, the network is scaled up overall, and a group feature analysis module is added. First, the image is extracted; the extracted regions are fed into the proposed network for training; the final recognition result is obtained based on the training results, greatly improving the recognition and classification accuracy of fingerprint activity detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of biometric identification applications, and in particular to a fingerprint activity detection method based on group feature analysis. Background Technology

[0002] Fingerprints are a commonly used biometric feature in identity authentication, and fingerprint authentication systems are widely used in finance, e-commerce, and other fields. However, like other biometric authentication systems such as facial and iris recognition, fingerprint authentication systems are susceptible to fraud attacks. Fingerprints can be replicated using materials (rubber, silicone, capacitive adhesive, etc.) to create fingerprint films that can successfully pass through fingerprint authentication systems, increasing the system's risk. The emergence of forged fingerprints has placed higher security demands on current fingerprint authentication systems. Forged fingerprints can launch various attacks against FLD (Fingerprint Liveness Detection) systems, necessitating the implementation of fingerprint liveness detection techniques to prevent such attacks. Fingerprint liveness detection is a component of fingerprint recognition systems used to detect the activity of input fingerprints and distinguish between live and forged fingerprints. In recent years, attacks on fingerprint liveness detection systems based on forged fingerprints have proliferated; therefore, an efficient fingerprint liveness detection system is needed to differentiate fingerprint images.

[0003] In recent years, fingerprint activity detection methods have emerged based on both hardware and software layers. Hardware-level methods enhance authentication system security by introducing devices capable of acquiring more biometric information and performing multiple verifications of the entered fingerprint. Software-level methods primarily utilize image detection technology to distinguish genuine from fake fingerprints after they have been entered. Compared to hardware detection methods, software detection methods are a more cost-effective approach based on image texture features and are easier to implement, thus attracting widespread attention. However, most current software-level algorithms are based on shallow hand-crafted feature design and SVM (Support Vector Machine) solutions, lacking in-depth mining of the texture features of fingerprint information.

[0004] Traditional methods rely on researchers' understanding of certain fingerprint features and employ conventional image processing techniques to extract these features for activity detection. One example is a texture feature algorithm combining LBP and wavelet transform. The LBP operator extracts texture features from the fingerprint image, while wavelet energy describes the frequency and direction information of the fingerprint image and its lines. In 2010, researchers proposed a heuristic algorithm for automatically extracting and locating pores. This algorithm uses two different types of filters to perform high-pass filtering on the spectrum of the fingertip image to obtain pore detection, and then associates the extracted pores with small circles to achieve pore localization.

[0005] Current fingerprint recognition technology research largely focuses on classification methods based on extracted features and feature lines, such as spirals, loops, and arcs, or on residual structures. These structures require standardized image-driven networks to automatically learn deep fingerprint features before making a final prediction of authenticity. These detection techniques have low network detection efficiency, require numerous practical parameters, lack precision in classifying forged fingerprints, have limited applicability, and rarely utilize the smoothness of the fingerprint image's lines to assess fingerprint activity and determine authenticity. Summary of the Invention

[0006] The main objective of this invention is to provide a fingerprint activity detection method based on group feature analysis, which solves the problems of low network detection efficiency and insufficient accuracy in identifying and classifying forged fingerprints in the prior art. By constructing a network model based on group features and studying a better group feature analysis method for judging fingerprint features extracted by the convolutional neural network model, the invention aims to determine the authenticity of fingerprints.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a fingerprint activity detection method based on group feature analysis, comprising the following steps: S1: Obtain the fingerprint dataset and perform image preprocessing; S2: Build a GAFA-Net detection model with multiple attention mechanisms; S3: Divide the dataset into a training set, train the model using the training set, and save the optimal model; S4: After performing the image preprocessing on the image to be detected, input it into the optimal model to classify the image and output the detection result.

[0008] In a preferred embodiment, the image preprocessing in S1 includes: S11: Input fingerprint image; S12: Extract ROI features from the image.

[0009] In the preferred embodiment, the model comprises nine stages, with the input and output terminals of each stage connected sequentially: The first stage contains a 3×3 convolutional layer; The second phase includes one layer of Fused-MBConv1; The third phase includes two layers of Fused-MBConv4; The fourth stage includes two layers of Fused-MBConv4; The fifth stage includes three layers of MBConv4; Phase 6 includes four layers of MBConv6; The seventh stage includes seven layers of MBConv6; The eighth stage includes the attention layer; The ninth stage consists of one 1×1 convolutional layer, adaptive average pooling, unfolded layer, and fully connected layer.

[0010] In the preferred embodiment, the attention layer is composed of an SK module, a Shuffle module, and a Triplet module connected sequentially, used for grouped feature analysis of the feature maps, wherein: The SK module processes the original input features using two different sizes of convolutional kernels to extract texture features of different dimensions; The Shuffle module groups features to focus on detailed image information and uses a dual-channel approach to obtain information from different feature maps. The Triplet module is used to capture cross-dimensional interactive feature information and enhance spatial feature fusion, processing feature information in three dimensions.

[0011] In a preferred embodiment, the SK module includes the following steps: S301: Perform convolution operations on the image to be tested using two different convolution kernels, and then perform batch normalization and BN layer processing on the convolutional feature maps in sequence. S302: Feature map obtained by fusing S301; S303: The fused feature map is processed by global flat pooling and fully connected layers to reduce dimensionality; S304: Differentiate the output features of S303 to obtain the differentiation vector; S305: Multiply the differentiation vector obtained in S304 by the range of the feature map processed in S301, and add the products to obtain the feature sum V.

[0012] In a preferred embodiment, the Shuffle module includes the following steps: S306: Divide the feature sum V into several feature maps according to the number of channel dimensions; S307: Each feature map group is divided into channel attention map branches and spatial attention map branches according to channels: S308: The feature maps of the channel attention map branches are subjected to global average pooling, which is then multiplied by the feature maps of the original channel attention map branches through a gating mechanism and activation function. S309: The feature maps of the spatial attention map branches are group normalized, then linearly calculated, and multiplied with the feature maps of the original spatial attention map branches; S310: Connect the two branches processed by S308 and S309, with the number of channels equal to the number of inputs, and aggregate them into a new feature map of the same size as the initial feature domain.

[0013] In a preferred embodiment, the Triplet module includes the following steps: S311: The feature map is processed in three branches: the first two branches are used to capture cross-dimensional interactive feature information between the channel attention map and the spatial attention map, and are rotated 90° counterclockwise along their corresponding axes; the last branch channel is used to construct the spatial attention mechanism. S312: Perform Z-pooling on the three branch feature maps after processing S311; S313: Convolution and batch normalization layers are applied to the feature maps processed by Z-pool; S314: Generate corresponding attention weights from the output feature maps of S313 using the Sigmoid function, and multiply them with the original feature domain; S315: The first two branches are rotated 90° clockwise along their respective axes; S316: The outputs of the three branches are averaged to obtain the final feature map.

[0014] This invention provides a fingerprint activity detection method based on group feature analysis, including acquiring a fingerprint dataset and performing image preprocessing; constructing a GAFA-Net (Group Feature Analysis Fusion Network) detection model; dividing the dataset into a training set, training the model using the training set, and saving the training set; and preprocessing the image to be detected and inputting it into the optimal model to classify the image. This invention is used to distinguish between forged and genuine fingerprints. Based on the EfficientNetV2-S model, the network is scaled up overall, and a group feature analysis module is added. This solves the technical problems of low network detection efficiency and insufficient accuracy in identifying and classifying forged fingerprints in existing technologies. First, the image is extracted, and the extracted regions are fed into the proposed network for training. After preprocessing the image to be detected, it is input into the optimal model, which classifies the image to obtain the final recognition result, greatly improving the recognition and classification accuracy of FLD (Finger Detection and Recognition). Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the detection method of the present invention; Figure 2 This is a flowchart of the data preprocessing process of the present invention; Figure 3 This is a schematic diagram of the process of extracting local blocks. Figure 4 This is a diagram of the detection network framework in this invention; Figure 5This is a schematic diagram of the SK attention mechanism processing flow used in this invention; Figure 6 This is a flowchart of the SK attention mechanism method used in this invention; Figure 7 This is a schematic diagram of the Shuffle attention mechanism processing flow in this invention; Figure 8 This is a flowchart of the Shuffle attention mechanism method in this invention; Figure 9 This is a schematic diagram of the Triplet attention mechanism processing flow in this invention; Figure 10 This is a flowchart of the Triplet attention mechanism method in this invention. Detailed Implementation

[0016] Example 1 like Figure 1-10 As shown, a fingerprint activity detection method based on group feature analysis includes the following steps: S1: Obtain the fingerprint dataset and perform image preprocessing.

[0017] S2: Build a GAFA-Net detection model with multiple attention mechanisms.

[0018] S3: Divide the dataset into a training set, train the model using the training set, and save the optimal model.

[0019] S4: After preprocessing the image to be detected, input it into the optimal model to classify the image and output the detection result.

[0020] In this embodiment, the detection model is named GAFA-Net (Group feature analysis fusion Network) detection model. It is based on the EfficientNetV2-S model, and the network has been scaled up as a whole, with a group feature analysis module added on top of it.

[0021] The data preprocessing in step S1 is to make the data meet the input model and be easy to process; that is, to modify the original data to make it easier to process.

[0022] First, the image dataset is preprocessed, and the processed images are divided into training and test sets. The training set is used to train the model and save it. The output results are judged to be true or false. The current recognition result is output according to the model, which improves the recognition and classification accuracy of the current FLD.

[0023] In the preferred embodiment, the image preprocessing in S1 includes: S11: Input fingerprint image.

[0024] S12: Extract ROI features from the data.

[0025] like Figure 2-3 As shown, in this embodiment, ROI (Region of Interest) feature extraction includes first calculating the mean of the fingerprint image. The fingerprint image is divided into several small blocks, and the standard deviation of each block is calculated. Compare the standard deviation of each piece with the mean. The image patches are stitched together to form the ROI output.

[0026] In step S12 of this embodiment, standard deviation is used, the threshold is the mean, and the selection condition is set to... Then, the data can be spliced ​​together; in practical applications, other data can also be used to achieve this.

[0027] In this embodiment, the processed image is input into the GAFA-Net model and trained.

[0028] In the preferred embodiment, the model consists of nine stages, with the input and output terminals of each stage connected sequentially.

[0029] The first stage consists of a 3×3 convolutional layer.

[0030] The second phase includes a layer of Fused-MBConv1.

[0031] The third stage includes two layers of Fused-MBConv4.

[0032] The fourth stage includes two layers of Fused-MBConv4.

[0033] The fifth stage includes three layers of MBConv4.

[0034] Phase 6 includes four layers of MBConv6.

[0035] The seventh stage includes seven layers of MBConv6.

[0036] The eighth stage includes the attention layer.

[0037] The ninth stage consists of one 1×1 convolutional layer, adaptive average pooling, unfolded layer, and fully connected layer.

[0038] Compared to the original EfficientNetV2-S network structure, the GAFA-Net model in this embodiment scales the original EfficientNetV2-S network structure proportionally, making the network more lightweight, improving the network detection speed, and adding an attention mechanism, which greatly improves the accuracy of identifying and classifying forged fingerprints.

[0039] Compared to existing technologies, the GAFA-Net network structure has fewer parameters, consumes less computational resources, and is more versatile for fingerprint activity detection. Even when faced with mismatches with the trained data, it can still maintain comparable performance.

[0040] like Figure 4 As shown, the GAFA-Net network is based on EfficientNetV2-S and includes Stem, Layer1, Layer2, Layer3, Layer4, Layer5, Layer6 and Head layers.

[0041] The main feature extraction is performed in the six convolutional layers at different scales. The first three layers, Layer 1, Layer 2, and Layer 3, use Fused-MBConv convolutions to perform convolutional residual connections on the input, extracting shallow and deep features of the image. The last three layers, Layer 4, Layer 5, and Layer 6, use MBConv convolutions to extract image features from a separable convolutional perspective and perform residual connections with the shallow features.

[0042] Existing technologies have proposed the idea of ​​convolutional layers, which, through stacking subsequent convolutional layers, can extract deep features that were not extracted in the previous layer. In this embodiment, the GAFA-Net network... Figure 1 The number of operation layers in Layer 1 to Layer 6 are 1, 2, 2, 3, 4, and 6 respectively. The original EfficientNetV2-S is scaled proportionally, and the number of convolutions in the middle six layers is reduced to half of the original number. This optimizes the model structure and reduces the number of parameters. Through heatmap experiments, the scaling is optimized for fingerprint activity detection tasks, thereby improving the efficiency and accuracy of fingerprint activity detection.

[0043] The extracted ROI image is input into the GAFA-Net model, and the EfficientNetV2-S is scaled proportionally to make the network more lightweight. The images are processed sequentially and a discrimination result is generated to identify whether the fingerprint is a live fingerprint or a forgery.

[0044] MBConv (Inverted Residual) is an inverted linear bottleneck layer with depthwise separable convolutions. In this embodiment, Fused MBConv (Fused Inverted Residual) is used to fuse the first and second convolutions into a single 3x3 convolution, solving the problem of slow depthwise convolutions. This makes MBConv faster and saves read / write time.

[0045] In this embodiment, the main body of the GAFA-Net network consists of stem, block, and head. The block includes a scaled EfficientNetV2-S model and an attention layer.

[0046] In the preferred scheme, the attention layer consists of an SK (Selective Kernel) module, a Shuffle module, and a Triplet module.

[0047] In this embodiment, three attention mechanisms were added to the network model to improve the accuracy of identifying and classifying forged fingerprints.

[0048] In this embodiment, SK improves its feature extraction capabilities by introducing selective multi-scale convolutional kernels and attention mechanisms. For example... Figure 5 As shown, the SK attention mechanism uses two different convolutional kernels to process the original input features, extracting the depth features of the original fingerprint image by generating statistical data features with different receptive field sizes. In fingerprint activity detection, fingerprint texture information is key to determining the activity of a fingerprint image. Multi-receptive field information interaction learning can effectively learn texture features of different dimensions, improving the network's recognition accuracy.

[0049] like Figure 7 As shown, Shuffle processes features in groups to better focus on the details of the image and uses a dual-channel approach to obtain information from different feature domains.

[0050] like Figure 9 As shown, the Triplet module is used to capture cross-dimensional interactive feature information and enhance spatial feature fusion. By processing information from three different dimensions, Triplet improves the robustness of the fingerprint activity detection algorithm.

[0051] In the preferred embodiment, the SK module includes the following steps: S301: Perform convolution operations on the image to be tested using two different convolution kernels, and then perform batch normalization and BN layer processing on the convolutional feature maps in sequence.

[0052] S302: Feature map obtained by fusing S301.

[0053] S303: The fused feature map is processed by global flat pooling and fully connected layers to reduce dimensionality.

[0054] S304: Differentiate the output features of S303 to obtain the differentiation vector.

[0055] S305: Multiply the differentiation vector obtained in S304 by the value range after segmentation in S301, and add the products to obtain the feature sum V.

[0056] like Figure 6 As shown, when the fingerprint feature map is processed by the SK module, the image generally goes through three stages: segmentation, fusion, and selection. During fusion, it needs to be fused with the shallow features after segmentation to learn different types of knowledge. Finally, the learned knowledge is selected to obtain the output feature map.

[0057] Step S301 is the segmentation stage.

[0058] In this embodiment, it is assumed that the feature domain of the original image is It first undergoes two types of convolution with kernels of 3 and 5 to obtain a new feature domain. , After convolution, the feature maps are processed sequentially through batch normalization and BN layers.

[0059] Step S302 is the fusion stage.

[0060] During the fusion phase, feature reduction values ​​are needed. These reduction values ​​are feature vectors obtained from the feature maps, which can adaptively adjust the main features in the feature domain. First, the feature domain obtained from S401 is fused: .

[0061] In step S303, the dimensionality of the fused feature map is reduced.

[0062] First, the results obtained from processing S302 are processed using global average pooling (gp) to obtain a single-channel statistical data S. After reducing the information in the H and W dimensions, the statistical data Sc for each channel can be represented by the following formula: ; Where gp represents global equal pooling.

[0063] Then, the Sc result is processed by a fully connected layer to obtain a more efficient reduction in dimension Z, as shown in the following formula: ; Where fc represents a fully connected layer. It is the ReLU function. It is batch normalization, W is a value belonging to The characteristic domain of space.

[0064] Let d be a scaled value, and the scaling factor be r, then: .

[0065] C is the number of channels, r is the attenuation rate, and L is to ensure that the value of d does not become a very small constant.

[0066] r is the attenuation rate, which can be set according to actual needs. The selection phase begins, including steps S304 and S305.

[0067] Step S304 involves differentiating the processing results after the segmentation and fusion stages, that is, differentiating the output feature Z of S303 into two different sets of feature values ​​a and b using Softmax, where each element... ai , bi It can be expressed by the following formula: ; ; in, η i and ξ i These are the i-th elements of vectors Az and Bz, respectively, where A and B both belong to... Domains, a and b are respectively and Each of their respective light attention vectors.

[0068] Step S305: Multiply the vectors a and b obtained in step S304 with the feature domains U1 and U2 in S301, and sum the products to obtain the feature sum V: ; Where V refers to all features, as shown below: .

[0069] In the preferred embodiment, the Shuffle module includes the following steps: S306: Divide the complete feature domain of V into several groups of feature maps according to the number of channel dimensions.

[0070] S307: Divide a set of feature maps into channel attention map branches and spatial attention map branches according to channels.

[0071] S308: The feature maps of the channel attention map branches are subjected to global average pooling, which is then processed through a gating mechanism and activation function, and multiplied with the feature maps of the original channel attention map branches.

[0072] S309: The feature maps of the spatial attention map branches are normalized using groups, linearly calculated, and multiplied with the feature maps of the original spatial attention map branches.

[0073] S310: Connect the two branches processed by S308 and S309, with the number of channels equal to the number of inputs, and aggregate them into a new feature map of the same size as the initial feature domain.

[0074] like Figure 8 As shown, the Shuffle module in this implementation groups the features, allowing attention to small details in the fingerprint image. For one group, a dual-channel approach is used to extract detailed features. The fingerprint image is segmented from a channel perspective, and the generated channel statistical feature data and spatial feature data are used to extract information from the original features. By concatenating the channel and spatial information, information from different feature domains of the fingerprint image can be obtained.

[0075] In step S306, the feature map output after processing in step S405 is divided into G groups according to the channel dimension, that is: .

[0076] In step S307, any set of feature maps from step S306 is selected and divided into two branches according to channels H and W, namely: .

[0077] The two branches are used to generate channel attention maps and spatial attention maps, respectively.

[0078] In step S308, on the path for generating the channel attention map: (1) Perform global average pooling on the feature domain to embed global information, as shown in the following formula: ; The obtained channel statistics S belong to Spatial domain.

[0079] (2) The result of global average pooling on the path of the generated channel attention map is multiplied by a gating mechanism and a sigmoid activation function, and the result is then multiplied with the original feature map to generate a compact feature. The formula is defined as follows: ; in, , used to scale and move parameters.

[0080] In step S309, on the path of generating the spatial attention map: (1) Spatial statistical information is obtained by group normalization of the input feature domain; (2) In terms of spatial information processing, the results of the path generated in the previous step of generating the spatial attention map are enhanced by linear calculation. The product of this product and the original feature is then taken, resulting in the following final output: ; in, Represents the Sigmoid operation. .

[0081] In step S310, the two output branches from steps S308 and S309 are connected, and the number of channels is equal to the number of inputs, as shown in the following formula: .

[0082] Each channel group is processed and connected in both spatial and channel dimensions, flowing along the channel dimension to aggregate into a new feature domain with the same size as the initial feature domain.

[0083] In the preferred embodiment, the Triplet module includes the following steps: S311: The feature map is processed in three branches: the first two branches are used to capture cross-dimensional interactive feature information between the channel attention map and the spatial attention map, and are rotated 90° counterclockwise along their corresponding axes; the last branch channel is used to construct the spatial attention mechanism.

[0084] S312: Perform Z-pooling on the three branch feature maps after processing S311.

[0085] S313: Convolution and batch normalization layers are applied to the feature maps processed by Z-pool.

[0086] S314: Generate corresponding attention weights from the output feature map of S313 using Sigmoid, and multiply them with the original feature domain.

[0087] S315: The first two branches are rotated 90° clockwise along their respective axes.

[0088] S316: The outputs of the three branches are averaged to obtain the final feature map.

[0089] The Triplet module uses a comprehensive analysis method based on H, W, and C dimensions to analyze the new features. It then uses convolution to generate the final statistical feature data, which is used to extract the key information from the input. Finally, the information from the three dimensions is summed to obtain the new features.

[0090] like Figure 10 As shown, the specific operation steps of the Triplet module are as follows: In step S311, the input feature map is divided into three branches for processing. Two of the three branches are used to capture cross-dimensional interactive feature information between channel C and spaces H and W. The last channel is used to construct a spatial attention mechanism to enhance spatial feature information.

[0091] For example, in this embodiment, in the first branch, the H dimension and the C dimension establish an interaction relationship. The input tensor x is then directed along... H The axis is rotated 90° counterclockwise. The tensor after rotation. Shape .

[0092] In the second branch, an interaction is established between the W dimension and the C dimension, with the input tensor... along W The axis is rotated 90° counterclockwise. The tensor after rotation. Shape .

[0093] In the third branch, spatial dimensions interact with each other.

[0094] In step S312, the results of the rotation of the first and second branches, as well as the feature map of the third branch, are processed. Z-pool The processing, where the result of Z-pool processing can be expressed by the following formula: .

[0095] After Z-pooling, each branch obtains a shape of... tensor , , .

[0096] The Z-pool layer is used to reduce the channel dimension, reducing the original dimension to 2-dimensional, and connecting the average pooling feature and the maximum pooling feature in this dimension, which preserves rich tensor information while reducing the amount of computation.

[0097] In step S313, the tensor , and By using convolution and batch normalization layers, a dimension of [value] is obtained. The intermediate output features.

[0098] In step S314, corresponding attention weights are generated for each tensor in step S313, and multiplied with the original feature domain to make them the same dimension as the original feature map.

[0099] Step S315: Rotate the results of the first branch and the second branch clockwise by 90° along the H and W axes respectively, so that they are consistent with the size of the input feature domain, while keeping the output of the third branch unchanged.

[0100] Step S316: Average the outputs of the three branches to obtain the following features: .

[0101] This invention discloses a fingerprint activity detection method based on deep learning, the specific process of which is as follows: Figure 5 As shown, the LivDet2011, LivDet2013, and LivDet2015 datasets were selected in this embodiment. The experimental data used in this embodiment are shown in Table 1.

[0102] Table 1

[0103] Figure 1 This diagram illustrates the process of extracting the ROI.

[0104] In this embodiment, FRR (False Rejection Rate) and FAR (False Acceptance Rate) are two main parameters used to evaluate the performance of fingerprint recognition algorithms.

[0105] In step S4, ACE (Average Classification Error) was used as the scoring criterion, and a good result was obtained. Therefore, this system can effectively detect forged fingerprints. The calculation method for ACE is shown in the following formula: .

[0106] In the above formula, FRR is the proportion of misclassifications in fake fingerprints, and FAR is the proportion of misclassifications in genuine fingerprints. The calculation formula is as follows: ; .

[0107] The experimental results are presented in Tables 2, 3, and 4. Table 2 shows the ACE evaluation of this method under the same material conditions; Table 3 shows the ACE evaluation across materials when the test set uses different forged fingerprint materials than the training set; and Table 4 shows the ACE values ​​of data acquired using different sensors in the training and datasets. In Tables 3 and 4, E stands for EcoFlex, G for Gelatine, L for Latex, S for Silgum, W for Wood Glue, and M for Modasil; I represents the Italdata sensor; B represents the Biometrika sensor; S represents the Sagem sensor; and H represents the GreenBit sensor. For example, B2011 represents the portion collected using the Biometrika sensor in the LivDet2011 dataset; E, G, and L represent portions collected using a specific sensor from a particular dataset, with materials of EcoFlex, Gelatine, and Latex.

[0108] Table 2

[0109] Table 3

[0110] Table 4

[0111] The values ​​in the table are percentages, representing the proportion of misclassified cases out of the total (average classification error rate). The closer the number is to 0, the better the model's classification performance.

[0112] The experimental results are shown in the table below, compared with existing technologies: Table 5

[0113] Table 6

[0114] Table 7

[0115] Table 5 shows the experimental results on the LivDet2011 dataset, Table 6 shows the experimental results on the LivDet2013 dataset, and Table 7 shows the experimental results on the LivDet2015 dataset.

[0116] Conclusion: On the LivDet2011 dataset, the performance improved by 0.92%, reaching state-of-the-art (SOTA). On the LivDet2013 dataset, the performance was average, ranking third. On the LivDet2015 dataset, the performance improved by 0.12%, also reaching SOTA.

[0117] Table 8

[0118] Note: Sensors: B-Biometrika, I-ItalData. Materials: E-EcoFlex, G-Gelatine, L-Latex, M-Modasil, W-WoodGlue, S-Silgum.

[0119] Table 9

[0120] Note: Sensors: H-Hi_Scan, G-GreenBit, S-Sagem Table 8 shows cross-material experiments, and Table 9 shows cross-sensor experiments; both experiments demonstrate significant improvements in recognition and classification accuracy.

[0121] In this embodiment, GAFA-Net is used to distinguish between forged and genuine fingerprints. Based on the EfficientNetV2-S model, the network has been scaled up overall, and a group feature analysis module has been added. First, Regions of Interest (ROIs) are extracted from the image; the extracted regions are then fed into the proposed network for training; the image to be detected is preprocessed and then input into the trained optimal network model to classify the image and output the detection results. Experiments have shown that the network proposed in this invention effectively improves the recognition and classification accuracy of current fingerprint activity identification.

[0122] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for detecting the activity of a fingerprint based on the analysis of packet characteristics, characterized in that, Includes the following steps: S1: Obtain the fingerprint dataset and perform image preprocessing; S2: Build a GAFA-Net detection model with multiple attention mechanisms; S3: Divide the dataset into a training set, train the model using the training set, and save the optimal model; S4: After performing the image preprocessing on the image to be detected, input the optimal model to classify the image and output the detection result; The GAFA-Net detection model consists of nine stages, with the input and output of each stage connected sequentially: The first stage contains a 3×3 convolutional layer; The second phase includes one layer of Fused-MBConv1; The third phase includes two layers of Fused-MBConv4; The fourth stage includes two layers of Fused-MBConv4; The fifth stage includes three layers of MBConv4; Phase 6 includes four layers of MBConv6; The seventh stage includes seven layers of MBConv6; The eighth stage includes the attention layer; The ninth stage includes one 1×1 convolutional layer, adaptive average pooling, unfolded layer, and fully connected layer. The attention layer consists of an SK module, a Shuffle module, and a Triplet module connected sequentially, used for grouped feature analysis of the feature maps, wherein: The SK module processes the original input features using two different sizes of convolutional kernels to extract texture features of different dimensions; The Shuffle module groups features to focus on detailed image information and uses a dual-channel approach to obtain information from different feature maps. The Triplet module is used to capture cross-dimensional interactive feature information and enhance spatial feature fusion, processing feature information in three dimensions.

2. The fingerprint activity detection method based on grouping feature analysis according to claim 1, characterized in that, The image preprocessing in S1 includes: S11: Input fingerprint image; S12: Extract ROI features from the image.

3. The fingerprint activity detection method based on grouping feature analysis according to claim 1, characterized in that, The SK module includes the following steps: S301: Perform convolution operations on the image to be tested using two different convolution kernels, and then perform batch normalization and BN layer processing on the convolutional feature maps in sequence. S302: Feature map obtained by fusing S301; S303: The fused feature map is processed by global average pooling and fully connected layers to reduce dimensionality; S304: Differentiate the output features of S303 to obtain the differentiation vector; S305: Multiply the differentiation vector obtained in S304 by the range of the feature map processed in S301, and add the products to obtain the feature sum V.

4. The fingerprint activity detection method based on grouping feature analysis according to claim 3, characterized in that, The Shuffle module includes the following steps: S306: Divide the feature sum V into several feature maps according to the number of channel dimensions; S307: Each feature map group is divided into channel attention map branches and spatial attention map branches according to channels: S308: The feature maps of the channel attention map branches are subjected to global average pooling, which is then multiplied by the feature maps of the original channel attention map branches through a gating mechanism and activation function. S309: The feature maps of the spatial attention map branches are group normalized, then linearly calculated, and multiplied with the feature maps of the original spatial attention map branches; S310: Connect the two branches processed by S308 and S309, with the number of channels equal to the number of inputs, and aggregate them into a new feature map of the same size as the initial feature domain.

5. The fingerprint activity detection method based on grouping feature analysis according to claim 3, characterized in that, The Triplet module includes the following steps: S311: The feature map is processed in three branches: the first two branches are used to capture cross-dimensional interactive feature information between the channel attention map and the spatial attention map, and are rotated 90° counterclockwise along their corresponding axes; the last branch channel is used to construct the spatial attention mechanism. S312: Perform Z-pooling on the three branch feature maps after processing S311; S313: Convolution and batch normalization layers are applied to the feature maps processed by Z-pool; S314: Generate corresponding attention weights from the output feature maps of S313 using the Sigmoid function, and multiply them with the original feature domain; S315: The first two branches are rotated 90° clockwise along their respective axes; S316: The outputs of the three branches are averaged to obtain the final feature map.