OCT fingerprint anti-counterfeiting method based on physiological structure consistency

Abstract

The application discloses an OCT fingerprint anti-counterfeiting method based on physiological structure consistency, which comprises the following steps: collecting OCT images of different fingerprints, extracting ROI to obtain ROI images, constituting a data set, dividing the data set into a training set and a verification set, and labeling part of the ROI images in the training set and the verification set; the OCT fingerprint anti-counterfeiting method based on physiological structure consistency identifies the OCT images of the fingerprints to be detected by constructing a double-branch model, obtains two intermediate features and two segmentation results, and realizes effective differentiation of real and fake fingerprints by calculating the similarity between the two intermediate features and the two segmentation results; in the second-stage training process, the second intermediate feature output by the second segmentation network is forced to align with the second intermediate feature output by the first segmentation network, and the constraint ensures that the double-branch model has highly similar feature expressions for real samples, so that the double-branch model can effectively detect known counterfeiting means.

Description

Technical Field

[0001] This invention belongs to the field of fingerprint anti-counterfeiting technology, specifically relating to an OCT fingerprint anti-counterfeiting method based on physiological structural consistency. Background Technology

[0002] Fingerprints are widely used as a biometric identification method due to their uniqueness and stability. Automated fingerprint recognition systems have been deployed in various applications such as access control and mobile authentication. However, these systems remain vulnerable to forgery attacks. Attackers can use materials such as silicone, gelatin, or 3D-printed replicas to forge fingerprints and deceive automated fingerprint recognition systems.

[0003] According to existing research, external fingerprints are determined by internal fingerprints formed in the dermis layer beneath the skin of the fingertip. Internal fingerprints can be used for both identification and defense against forgery attacks. This is because real human skin has a complex physiological structure, containing the epidermis, dermis, and sweat glands. Due to this complexity, it is difficult for forgery samples to completely and accurately reproduce the skin structure. This leads to forgery samples and real skin tissue potentially having identical external fingerprints, while their internal structures and fingerprints differ significantly. Accurately identifying these differences is crucial for automated fingerprint recognition systems to resist forgery attacks. Furthermore, when external fingerprints are damaged due to injury or labor, they may be misjudged by automated fingerprint recognition systems; combining this with internal fingerprint analysis helps reduce the error rate. Optical coherence tomography (OCT) is currently the primary technique for acquiring internal fingerprints. It can acquire tissue structures at a certain depth beneath the finger's skin and reconstruct them into three-dimensional data, providing a new method for automated fingerprint recognition systems to defend against forgery attacks.

[0004] Existing OCT fingerprint anti-counterfeiting technologies are divided into manual feature extraction and deep learning methods. Manual feature extraction methods primarily involve manually designing indicators to extract common features from real samples; samples matching these features are considered genuine. However, fixed algorithms struggle to adapt to individual differences, and feature extraction design relies on the designer's understanding of the fingertip OCT image. Deep learning methods rely on neural networks to automatically find differences between real and counterfeit OCT images, requiring the neural network to abstractly model the OCT images of real samples. While introducing neural networks greatly enhances generalization, it may overfit to features in the OCT image unrelated to authenticity, leading to a higher false positive rate. Summary of the Invention

[0005] The purpose of this invention is to address the problems raised in the background art by proposing an OCT fingerprint anti-counterfeiting method based on the consistency of physiological structure.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention proposes an OCT fingerprint anti-counterfeiting method based on physiological structural consistency, comprising:

[0008] OCT images of different fingerprints were collected, and ROI images were extracted to form a dataset. The dataset was divided into a training set and a validation set, and a portion of the ROI images in the training set and the validation set were labeled.

[0009] A dual-branch model is constructed and trained in stages. The dual-branch model includes a first segmentation network and a second segmentation network with the same structure. Each segmentation network includes an extractor and a reconstructor connected in sequence. The extractor outputs intermediate features, and the reconstructor outputs segmentation results about the physiological structure of the skin. The segmentation results include segmented images of four categories: background, stratum corneum, sweat glands and active epidermal junction layer.

[0010] During the training of the dual-branch model, in the first stage, the first segmentation network is trained under full supervision using a labeled training set to obtain the trained first segmentation network.

[0011] In the second stage, the first step is to train the extractor in the second segmentation network using an unlabeled training set, and the second step is to train the extractor and reconstructor of the second segmentation network using a labeled training set. The first and second steps are executed alternately to obtain the trained second segmentation network.

[0012] The OCT image of the fingerprint to be detected is acquired, and after the preset ROI image is extracted, it is input into the trained first segmentation network and second segmentation network respectively to obtain two first intermediate features and two first segmentation results. The anti-counterfeiting score is calculated, and the anti-counterfeiting score is used to determine whether the OCT image to be detected comes from a real finger, thereby realizing fingerprint anti-counterfeiting.

[0013] Preferably, the process of acquiring the ROI image includes:

[0014] Obtain OCT scan data of the fingerprint region of the finger to obtain the fingerprint OCT image. Calculate the average pixel brightness of each row of the OCT image and select the row with the highest brightness. Line, using a width of The sliding window is for the brightest area. The rows are cropped according to a preset overlap rate, and the resulting image after each cropping is the [image type / description]. The ROI image.

[0015] Preferably, OCT images of different fingerprints are acquired, and ROI images are obtained by ROI extraction. Then, each ROI image is stretched and deformed, and finally, an interpolation algorithm is used to adjust it to the desired shape. The dataset consists of all ROI images with the specified resolution and adjusted by the interpolation algorithm.

[0016] The annotation process includes: annotating the pixels in the ROI image that belong to the background, stratum corneum, sweat glands, and active epidermal junction layer, respectively.

[0017] Preferably, the extractor includes a convolutional block, a first downsampling module, a second downsampling module, and a third downsampling module connected in sequence;

[0018] The convolutional block and each downsampling module include a convolutional layer, a batch normalization layer and an activation layer connected in sequence;

[0019] The reconstructor includes a first upsampling module, a second upsampling module, a third upsampling module, and an output prediction block connected in sequence;

[0020] Each of the upsampling modules includes a transposed convolutional layer, a batch normalization layer, and an activation layer connected in sequence, and the output prediction block includes a convolutional layer, an activation layer, and a binarization process connected in sequence.

[0021] Preferably, the step of performing fully supervised training on the first segmentation network using a labeled training set to obtain a trained first segmentation network includes:

[0022] Before training, the number of pixels belonging to the background, stratum corneum, sweat glands and active epidermal connective layer in the labeled training set is counted. Initial weights are set for any category among the background, stratum corneum, sweat glands and active epidermal connective layer to obtain the initial weights of each category. The initial weights are used as the weights of each category in the first round of training.

[0023] During a round of training:

[0024] The first segmentation network is forward propagated using a labeled training set to obtain the first predicted segmentation image of each sample in the labeled training set corresponding to the four categories.

[0025] Perform a first normalization on the weights of each category in the current round;

[0026] The first loss function is calculated based on the first predicted segmented image of the four categories, the real segmented image, and the weights of the four categories after normalization in the current round. The parameters of the first segmentation network are then updated through backpropagation.

[0027] The performance of the first segmentation network was evaluated using a validation set, and the first Dice coefficients for the four categories were calculated using the validation set during the evaluation process.

[0028] Calculate the first suggested weight for each category based on the first Dice coefficient for each category, and update the first normalized weights for each category in the current round based on the first suggested weights for each category. The updated weights for each category will then be used as the weights for each category in the next round of training.

[0029] Enter the next round of training, and continue the cycle until the first Dice coefficients of each category converge, thus obtaining the trained first segmentation network.

[0030] Preferably, when the first step and the second step are executed alternately, each execution of the first step and the second step in sequence is called a round. For each round:

[0031] The first step is to input each sample in the unlabeled training set into the extractors of the second segmentation network and the trained first segmentation network respectively to obtain two second intermediate features;

[0032] The difference between the two second intermediate features is calculated using the second loss function, and the parameters of the extractor in the second segmentation network are updated by backpropagation.

[0033] The second step is to perform forward propagation of the labeled training set on the second segmentation network to obtain the second predicted segmentation image corresponding to the four categories for each sample in the labeled training set.

[0034] The weights of each category in the current round are normalized a second time, while the weights of each category in the first round use the initial weights.

[0035] The third loss function is calculated based on the second predicted segmentation image of four categories, the real segmentation image, and the weights of the four categories after second normalization in the current round, and the parameters of the second segmentation network are updated by backpropagation.

[0036] The performance of the second segmentation network was evaluated using a validation set, and the second Dice coefficients for the four categories were calculated using the validation set during the evaluation process.

[0037] The second suggested weights for each category are calculated based on the second Dice coefficients for each category. The second suggested weights for each category are then updated based on the second suggested weights for each category in the current round. The updated weights for each category are then used as the weights for each category in the next round.

[0038] Enter the next round, and continue the loop until the second loss function and the second Dice coefficients of each category converge, resulting in the trained second segmentation network.

[0039] Preferably, when performing preset ROI image extraction on the OCT image of the fingerprint to be detected, a width of [missing information] is used. The sliding window is the brightest part of the OCT image of the fingerprint to be detected. Extracting the center region of the row yields a size of... The preset ROI images are used as inputs to the trained first and second segmentation networks, respectively.

[0040] Preferably, when calculating the anti-counterfeiting score, the cosine distance between two first intermediate features is first calculated, and then a third normalization is performed to obtain the intermediate feature similarity score.

[0041] Then, the third Dice coefficients are calculated between the segmented images of the two stratum corneum, the segmented images of the two sweat glands, and the segmented images of the two active epidermal junction layers in the two first segmentation results. Three third Dice coefficients are obtained, and the three third Dice coefficients are normalized in the fourth way to obtain the segmentation similarity score.

[0042] The anti-counterfeiting score is obtained by multiplying the intermediate feature similarity score and the segmentation similarity score.

[0043] Preferably, the step of determining whether the OCT image to be detected comes from a real finger based on the anti-counterfeiting score includes:

[0044] The anti-counterfeiting score is compared with a preset value. If the score is less than the preset value, the OCT image to be detected is from a real finger; otherwise, the OCT image to be detected is not from a real finger.

[0045] Preferably, the first and third loss functions both adopt the dynamic weighted cross-entropy loss function, and the second loss function adopts the cosine distance loss and L1 distance loss.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. This OCT fingerprint anti-counterfeiting method based on physiological structure consistency constructs a dual-branch model to identify the OCT image of the fingerprint to be detected, obtains two intermediate features and two segmentation results, and achieves effective differentiation between genuine and counterfeit fingerprints by calculating the similarity between the two intermediate features and the two segmentation results.

[0048] 2. In the second stage of training, the OCT fingerprint anti-counterfeiting method based on physiological structure consistency forces the second intermediate feature output by the second segmentation network to align with the second intermediate feature output by the first segmentation network. This constraint ensures that the dual-branch model has highly similar feature representations to real samples, enabling the dual-branch model to not only effectively detect known forgery methods, but also to achieve robust defense against unknown and novel forgery attacks by keenly capturing abnormal situations that deviate from the distribution of real physiological structures.

[0049] 3. The first and second loss functions of this method both adopt the dynamic weighted cross-entropy loss function. The weights are dynamically adjusted according to the Dice coefficient on the validation set in each round, giving higher suggestion weights to the target category with poor segmentation performance, and forcing the dual-branch model to pay equal attention to the physiological structure of all categories. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the OCT fingerprint anti-counterfeiting method based on physiological structural consistency of the present invention.

[0051] Figure 2 This is a structural block diagram of the internal and external fingerprint extraction network of the present invention;

[0052] Figure 3 This is a structural annotation diagram of the stratum corneum, sweat glands, and active epidermal junction layer in the ROI image of this invention;

[0053] Figure 4 This diagram illustrates the consistency of the dual-branch model of the present invention for segmenting real finger OCT images. Green represents the consistent segmentation results, and red represents the differing results.

[0054] Figure 5 This diagram illustrates the consistency of the segmentation results of the dual-branch model of the present invention for the OCT images of forged samples. Green represents the regions with consistent segmentation results, and red represents the regions with differences. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0057] like Figures 1-5 As shown, an OCT fingerprint anti-counterfeiting method based on physiological structural consistency is provided, including:

[0058] Step 1: Collect OCT images of different fingerprints and extract ROIs to form a dataset. Divide the dataset into a training set and a validation set, and label a portion of the ROI images in the training set and the validation set. In this embodiment, the training set contains 120 labeled images and 16,000 unlabeled images; the validation set contains 24 images.

[0059] Step 1.1, the process of extracting ROI images and constructing the dataset, includes:

[0060] OCT scan data of the fingerprint area of ​​the finger are acquired to obtain an OCT image of the fingerprint (one OCT image can represent a longitudinal section of the fingertip skin at a specific location). The average pixel brightness of each row of the OCT image is calculated, and the brightest pixel is selected. Line, using a width of A sliding window (the height of which can be 256 rows) is used for the window with the highest brightness. The rows are cropped according to a preset overlap rate, and the resulting image after each cropping is the [image type / description]. ROI image;

[0061] Considering that skin tissue is an elastic biological tissue, different degrees of pressure during the acquisition process can cause deformation of the skin tissue. Stretching and deformation operations were performed on each ROI image in both the horizontal and vertical directions (in the horizontal direction, a region of random width was cropped and scaled back). In the column, simulate horizontal stretching; in the vertical direction, locate the upper and lower boundaries of the skin according to the label, and randomly crop and scale it vertically. Okay, simulate tissue compression under different pressure levels, and then use an interpolation algorithm to adjust to... The resolution is high (the purpose is to simulate the deformation of skin tissue when a finger is pressed, and to improve the robustness of the neural network to deformation), and all ROI images adjusted by the interpolation algorithm constitute the dataset (that is, one ROI image adjusted by the interpolation algorithm is a sample in the dataset).

[0062] Step 1.2: Divide the dataset into a training set and a validation set, and label a portion of the ROI images in the training set (this portion is called the labeled training set), while leaving the other portion unlabeled (this portion is called the unlabeled training set), and label the ROI images in the validation set.

[0063] The annotation process includes: annotating the pixels in the ROI image that belong to the background, stratum corneum, sweat glands, and active epidermal junction layer, respectively.

[0064] Step 2: Construct a dual-branch model and train it in stages. The dual-branch model includes a first segmentation network and a second segmentation network with the same structure. Each segmentation network includes an extractor and a reconstructor connected in sequence. The samples in the dataset are used as the input of the extractor, and the extractor outputs intermediate features. The reconstructor outputs the segmentation results about the physiological structure of the skin. The segmentation results include segmented images of four categories: background, stratum corneum, sweat glands and active epidermal junction layer.

[0065] The extractor and reconstructor can both be built based on convolutional neural networks or both based on the Transformer architecture. When built based on convolutional neural networks (the extractor receives a single-channel grayscale ROI image with a resolution of 256×256 and outputs intermediate features with dimensions of (32, 32, 256), and the reconstructor receives intermediate features with dimensions of (32, 32, 256) and outputs physiological structure segmentation images with dimensions of (256, 256, 4), namely, segmentation images of four categories: background, stratum corneum, sweat glands, and active epidermal junction layer):

[0066] The extractor comprises a convolutional block (receiving a single-channel grayscale ROI image with a resolution of 256×256 and an output dimension of (256, 256, 32)), a first downsampling module (output dimension of (128, 128, 64)), a second downsampling module (output dimension of (64, 64, 128)), and a third downsampling module (output dimension of (32, 32, 256)) connected in sequence.

[0067] Each convolutional block and downsampling module includes sequentially connected convolutional layers (where the convolutional layer in the convolutional block has a kernel size of 3×3, stride of 1, padding of 1, channel depth of 1, and number of channels 32; the convolutional layer in the first downsampling module has a kernel size of 3×3, stride of 2, padding of 1, channel depth of 32, and number of channels 64; the convolutional layer in the second downsampling module has a kernel size of 3×3, stride of 2, padding of 1, channel depth of 64, and number of channels 128; and the convolutional layer in the third downsampling module has a kernel size of 3×3, stride of 2, padding of 1, channel depth of 128, and number of channels 256), a batch normalization layer, and an activation layer (ReLU activation function).

[0068] The reconstructor consists of a first upsampling module (output dimension (64, 64, 128)), a second upsampling module (output dimension (128, 128, 64)), a third upsampling module (output dimension (256, 256, 32)), and an output prediction block (output dimension (256, 256, 4)) connected in sequence.

[0069] Each upsampling module includes a transposed convolutional layer connected in sequence (the first upsampling module has a 4×4 kernel size, a stride of 2, a padding of 1, a channel depth of 256, and 128 channels; the second upsampling module has a 4×4 kernel size, a stride of 2, a padding of 1, a channel depth of 128, and 64 channels; the third upsampling module has a 4×4 kernel size, a stride of 2, a padding of 1, a channel depth of 64, and 32 channels). The output prediction block consists of a batch normalization layer and an activation layer (ReLU activation function). The output prediction block includes a convolutional layer (the kernel size of the convolutional layer is 1×1, stride is 1, padding is 0, channel depth is 32, and number of channels is 4) connected in sequence, an activation layer (Softmax activation function), and a binarization process (the maximum value of each pixel in the physiological structure probability distribution map is filtered, the channel with the highest probability value is set to 1, and the other channels are set to 0, thus generating a physiological structure segmentation image where each channel is a binary image).

[0070] When constructing a physiological structure segmentation image based on the Transformer architecture (the extractor receives a single-channel grayscale ROI image with a resolution of 256×256 and outputs intermediate features with dimensions of (256, 512); the reconstructor receives intermediate features with dimensions of (256, 512) and outputs segmentation images of physiological structures with dimensions of (256, 256, 4), namely, segmentation images of four categories: background, stratum corneum, sweat glands, and living epidermal junction layer):

[0071] The extractor consists of a tile embedding module (output dimension (256, 512)) and a Transformer encoding module (output dimension (256, 512)) connected in sequence. The tile embedding module includes a convolutional layer (with a kernel size of 16×16, stride of 16, padding of 0, channel depth of 1, and number of channels of 512) and a layer normalization layer connected in sequence. The Transformer encoding module contains 12 encoding blocks connected in sequence. Each encoding block includes a layer normalization layer, a multi-head self-attention layer (with 8 heads, each head having a feature dimension of 64, for a total dimension of 512), a layer normalization layer, a multi-layer perceptron layer (including two fully connected layers connected in sequence; the first fully connected layer expands the feature dimension from 512 to 2048, and the second fully connected layer compresses the feature dimension from 2048 back to 512), and an activation layer (GELU activation function).

[0072] The reconstructor comprises a Transformer decoding module (output dimension (256, 512)), a tile restoration module (output dimension (256, 256, 32)), and an output prediction block (output dimension (256, 256, 4)) connected in sequence. The Transformer decoding module consists of 12 sequentially connected decoding blocks. Each block includes a layer normalization layer, a multi-head self-attention layer (with 8 heads, each with a feature dimension of 64), a layer normalization layer, a multi-layer perceptron layer (comprising two sequentially connected fully connected layers; the first fully connected layer expands the feature dimension from 512 to 2048, and the second fully connected layer compresses the feature dimension back to 512), and an activation layer (GELU activation function). The tile restoration module includes a sequence reshaping layer (which spatially rearranges the output dimension (256, 512) of the Transformer decoding module to form a dimension (16, 16, 4)). The structure of the output prediction block is consistent with that of the output prediction block in the reconstructor built based on the convolutional neural network, and will not be described again. The transposed convolutional layer (with a kernel size of 16×16, a stride of 16, padding of 0, a channel depth of 512, and a number of 32), the batch normalization layer, and the activation layer (ReLU activation function) are also described.

[0073] The sequential connections described in this application all represent the direction from data input to output.

[0074] Step 3: During the training of the dual-branch model, in the first stage, the first segmentation network is trained in a fully supervised manner using the labeled training set to obtain the trained first segmentation network.

[0075] The second stage consists of two steps: First, training the extractor in the second segmentation network using an unlabeled training set, ensuring that the output of the extractor in the second segmentation network matches the output of the extractor in the trained first segmentation network. Second, training both the extractor and reconstructor of the second segmentation network using a labeled training set. The first and second steps are performed alternately to obtain the trained second segmentation network, which includes:

[0076] The dual-branch model is trained in two stages:

[0077] Step 3.1, First stage: The first segmentation network is trained under full supervision using a labeled training set (in this embodiment, the full supervision training has 200 epochs, a batch size of 8, and a learning rate of 0.0001), resulting in the trained first segmentation network, including:

[0078] Step 3.1.1: Before training, count the number of pixels in the labeled training set belonging to the background, stratum corneum, sweat glands, and active epidermal connective layer, respectively. Set initial weights for each of these categories, and calculate the initial weights using the following formula:

[0079] ;

[0080] in, For category The initial weights and categories It can be any one of the following categories: background, stratum corneum, sweat glands, and active epidermal junction. To belong to the category in the labeled training set The proportion of pixels (i.e., the categories in the labeled training set) The ratio of the number of pixels to the total number of pixels in the labeled training set.

[0081] Then, the initial weights for each category are obtained (i.e., the initial weights for the background, the initial weights for the stratum corneum, the initial weights for the sweat glands, and the initial weights for the active epidermal junction layer), and these initial weights serve as the weights for each category in the first round of training.

[0082] Step 3.1.2: Train the first segmentation network in multiple rounds, with each round containing multiple batches (i.e., dividing the labeled training set into multiple batches). For each batch, execute steps 3.1.2.1-3.1.2.3:

[0083] Step 3.1.2.1: Perform forward propagation on the first segmentation network using the labeled training set to obtain the first predicted segmentation images of the four categories corresponding to each sample in the labeled training set (i.e., obtain the first predicted segmentation results, and the first predicted segmentation results include the first predicted segmentation images of the four categories (i.e., the first predicted segmentation image of the background, the first predicted segmentation image of the stratum corneum, the first predicted segmentation image of the sweat glands, and the first predicted segmentation image of the active epidermal junction layer)).

[0084] Step 3.1.2.2: Perform the first normalization on the weights of each category in the current round;

[0085] The formula corresponding to the first normalization is as follows:

[0086] ;

[0087] in, For the current number Round Category The first normalized weights, For the current number Round Category The weights (before the first normalization). Represented by the serial number of each category, For the current number Round Category The weight, The total number of categories (in this embodiment, it is 4, namely, background, stratum corneum, sweat glands and active epidermal junction).

[0088] Step 3.1.2.3: Calculate the first loss function (using the dynamic weighted cross-entropy loss function) based on the first predicted segmentation image of the four categories, the real segmentation image, and the first normalized weights of the four categories in the current round, and update the parameters of the first segmentation network through backpropagation;

[0089] Wherein, the first loss function The calculation formula is as follows:

[0090] ;

[0091] in, This represents the total number of pixels in all samples during the current batch training process. for The Middle Each pixel belongs to the category binary labels, for The Middle Each pixel belongs to the category The predicted probability.

[0092] Continue iterating through batches until all batches in the current training round have been traversed, then complete the current training round and proceed to step 3.1.2.4;

[0093] Step 3.1.2.4: Use the validation set to evaluate the performance of the first segmentation network, and use the validation set to calculate the first Dice coefficients for the four categories during the evaluation process;

[0094] The formula for calculating the first Dice coefficient is as follows:

[0095] ;

[0096] in, For category The first Dice coefficient, The first segmentation network predicts the category based on the validation set. The set of pixels, To verify the set of categories The set of pixel real labels, In order to find the intersection, To find the number of pixels in the set, and The value is between 0 and 1, and the closer the value is to 1, the more accurate the output of the first segmentation network is.

[0097] Step 3.1.2.5: Calculate the first suggested weight of each category based on the first Dice coefficient of each category, and update the first normalized weight of each category in the current round based on the first suggested weight of each category. The updated weight of each category is used as the weight of each category in the next round of training.

[0098] The formula for calculating the weight of the first recommendation is as follows:

[0099] ;

[0100] in, For category The first recommended weight, This is a preset value (a small constant to prevent division by zero);

[0101] Specifically, the first normalized weights of each category in the current round are updated based on the first suggested weights of each category, as shown in the following formula:

[0102] ;

[0103] in, For the current number Round Category The updated weights, and used as the class in the next training round. The weight, This is the momentum coefficient (e.g., 0.2).

[0104] Step 3.1.2.6: Enter the next round of training and continuously loop (i.e. repeat steps 3.1.2.1-3.1.2.6) until the first Dice coefficient of each category converges (in this embodiment, convergence occurs when the first Dice of each category does not increase for 10 consecutive rounds), thus obtaining the trained first segmentation network.

[0105] Step 3.2, Second Stage:

[0106] In this embodiment, the second phase has 150 rounds, a batch size of 16 (each step in each round contains multiple batches), and a learning rate of 1e-5.

[0107] When the first and second steps are executed alternately, each execution of the first and second steps in sequence is called a round. For each round:

[0108] Step 3.2.1, First step:

[0109] Step 3.2.1.1: Freeze the parameters of the trained first segmentation network, and input each sample in the unlabeled training set into the extractor of the second segmentation network and the extractor of the trained first segmentation network respectively to obtain two second intermediate features;

[0110] Step 3.2.1.2: Use the second loss function (using cosine distance loss and L1 distance loss) to calculate the difference between the two second intermediate features, and backpropagate to update the parameters of the extractor in the second segmentation network. By minimizing the second loss function, the second intermediate features of the second segmentation network are aligned with the second intermediate features of the trained first segmentation network.

[0111] The second loss function The calculation formula is as follows:

[0112] ;

[0113] in,

[0114] ;

[0115] ;

[0116] in, and These are the cosine distance loss and the L1 distance loss, respectively. and These represent the outputs of the extractors from the first and second segmentation networks, respectively. It is the L2 norm. This is a preset value (a small constant to prevent division by zero). express The One element, express The One element, Dimensions are the outputs of each extractor.

[0117] Step 3.2.2, Second Step:

[0118] Step 3.2.2.1: Perform forward propagation of the labeled training set on the second segmentation network to obtain the second predicted segmentation images of the four categories corresponding to each sample in the labeled training set (i.e., obtain the second predicted segmentation results, and the second predicted segmentation results include the second predicted segmentation images of the four categories (i.e., the second predicted segmentation image of the background, the second predicted segmentation image of the stratum corneum, the second predicted segmentation image of the sweat glands, and the second predicted segmentation image of the active epidermal junction layer)).

[0119] Step 3.2.2.2: Perform a second normalization on the weights of each category in the current round, while using the initial weights for each category in the first round;

[0120] The formula corresponding to the second normalization is as follows:

[0121] ;

[0122] in, For the current number Round category The second normalized weights, For the current number Round Category The weights (before the second normalization). For the current number Round category The weight.

[0123] Step 3.2.2.3: Calculate the third loss function (using dynamic weighted cross-entropy loss function) based on the second predicted segmentation image of the four categories, the real segmentation image, and the weights of the four categories after second normalization in the current round, and update the parameters of the second segmentation network through backpropagation;

[0124] Among them, the third loss function The calculation formula is as follows:

[0125] ;

[0126] in, This represents the total number of pixels in labeled samples during the current training round. for The Middle Each pixel belongs to the category binary labels, for The Middle Each pixel belongs to the category The predicted probability.

[0127] Step 3.2.2.4: Use the validation set to evaluate the performance of the second segmentation network, and use the validation set to calculate the second Dice coefficients for the four categories during the evaluation process;

[0128] The formula for calculating the second Dice coefficient is as follows:

[0129] ;

[0130] in, For category The second Dice coefficient, The second segmentation network predicts the category based on the validation set. The set of pixels, To verify the set of categories The set of real pixel labels, and The value is between 0 and 1, and the closer the value is to 1, the more accurate the output of the second segmentation network.

[0131] Step 3.2.2.5: Calculate the second suggested weight of each category based on the second Dice coefficient of each category, and update the second normalized weight of each category in the current round based on the second suggested weight of each category. The updated weight of each category shall be used as the weight of each category in the next round.

[0132] The formula for calculating the weight of the second recommendation is as follows:

[0133] ;

[0134] in, For category The second recommended weight;

[0135] Specifically, the second-normalized weights of each category in the current round are updated based on the second-recommended weights of each category, as shown in the following formula:

[0136] ;

[0137] in, For the current number Round category The updated weights, and used as the category in the next round. The weight, The momentum coefficient is set to 0.2. An exponential moving average strategy is introduced to smoothly update the weights of each category, preventing model oscillations caused by sudden weight changes and effectively ensuring the stability of the training process.

[0138] Step 3.2.2.6: Proceed to the next round and continuously loop (i.e., continuously alternate between the first step and the second step, i.e., repeat steps 3.2.1.1-3.2.2.6) until the second loss function and the second Dice coefficients of each category converge (in this embodiment, the second loss function converges when it does not decrease for 10 consecutive rounds, and the second Dice coefficients of each category converge when they do not increase for 10 consecutive rounds), thus obtaining the trained second segmentation network.

[0139] Step 4: Obtain the OCT image of the fingerprint to be detected, and after extracting the preset ROI image, input it into the trained first segmentation network and second segmentation network respectively to obtain two first intermediate features and two first segmentation results. Calculate the anti-counterfeiting score, and determine whether the OCT image to be detected comes from a real finger based on the anti-counterfeiting score, thereby realizing fingerprint anti-counterfeiting.

[0140] Step 4.1: Following Step 1.1, perform preset ROI image extraction on the OCT image of the fingerprint to be detected: calculate the average pixel brightness of each row of the OCT image, and select the row with the highest brightness. Line, using a width of The sliding window (in this implementation, the height of the sliding window is set to 256 rows) is used to select the window with the highest brightness. Extract the center region of the row to obtain The intercepted region, and The cropped area is used as the preset ROI image.

[0141] Step 4.2: The preset ROI images are used as inputs to the trained first segmentation network and the second segmentation network, respectively. The extractors in each trained segmentation network output the first intermediate features, and the first intermediate features are used as inputs to the reconstructors in each trained segmentation network. The reconstructors output the first segmentation results (which include segmented images of four categories: background, stratum corneum, sweat glands and active epidermal junction layer), thus obtaining two first intermediate features and two first segmentation results.

[0142] Step 4.3: Based on the two first intermediate features and the two first segmentation results, calculate the anti-counterfeiting score, including:

[0143] Step 4.3.1: First, calculate the cosine distance between the two first intermediate features (i.e., the formula is...). The system then performs a third normalization (which is a truncated min-max normalization based on the statistical distribution to eliminate the influence of outliers on the score distribution) to obtain the intermediate feature similarity score. The formula is expressed as follows:

[0144] ;

[0145] in, For the third normalization, and These are the first intermediate features corresponding to the first trained segmentation network and the second intermediate features corresponding to the second trained segmentation network, respectively. The numerical range is between 0 and 1. The closer it is to 1, the more it indicates and The greater the difference, the more likely it is a forged fingerprint.

[0146] Step 4.3.2: Then calculate the third Dice coefficient (i.e., the formula is) between the segmented images of the two stratum corneum layers, the segmented images of the two sweat glands, and the segmented images of the two active epidermal junction layers in the two first segmentation results. The three third Dice coefficients are obtained, and a fourth normalization is performed on them (the fourth normalization is a truncated maximum-min normalization based on the statistical distribution to eliminate the influence of outliers on the score distribution), to obtain the segmentation similarity score; where the segmentation similarity score The formula is expressed as follows:

[0147] ;

[0148] in, For the fourth normalization, and These represent the outputs of the first segmentation network after training, for each category. The segmented image and the output of the trained second segmentation network are targeted at the category Segmented image, category It can be any one of the following categories: stratum corneum, sweat glands, and active epidermal junction. The numerical range is between 0 and 1. The lower the value, the more the segmentation results of the trained first segmentation network and the trained second segmentation network overlap, which conforms to the anatomical structure of real skin.

[0149] Step 4.3.3: Multiply the intermediate feature similarity score and the segmentation similarity score to obtain the anti-counterfeiting score. The formula is as follows:

[0150] .

[0151] Step 4.3.4: Determine whether the OCT image to be detected comes from a real finger based on the anti-counterfeiting score, including:

[0152] The anti-counterfeiting score is compared with a preset value. If it is less than the preset value, the OCT image to be detected is from a real finger; otherwise, the OCT image to be detected is not from a real finger and is a forged fingerprint.

[0153] To demonstrate the fingerprint anti-counterfeiting effectiveness of this method, it was compared with existing methods on three error rate metrics: EER (Equal Error Rate), BPCER10 (the error rate of real person presentation when the counterfeit presentation error rate is 10%), and BPCER20 (the error rate of real person presentation when the counterfeit presentation error rate is 5%). The existing methods are OCPAD (from the existing literature Liu, F., Liu, H., Zhang, W., Liu, G., & Shen, L. (2021). One-Class Fingerprint Presentation Attack Detection Using Auto-EncoderNetwork. IEEE Transactions on Image Processing, 30, 2394–2407.) and FFDPAD (from the existing literature Zhang, W., Liu, H., & Liu, F. (2021). Fingerprint Presentation Attack Detection by Learning in Frequency Domain. 2021 IEEE 2nd International Conference on Pattern Recognition and Machine Learning (PRML). The comparison data between 183–189. and MemAE (from existing literature Gong, D., Liu, L., Le, V., Saha, B., Mansour, MR, Venkatesh, S., & Van Den Hengel, A. (2019). Memorizing Normality to DetectAnomaly: Memory-Augmented Deep Autoencoder for Unsupervised AnomalyDetection. 2019 IEEE / CVF International Conference on Computer Vision (ICCV), 1705–1714.) is shown in Table 1.

[0154] Table 1

[0155]

[0156] As shown in Table 1, compared with existing methods, the error rate of this method is much lower in all three indicators, so the fingerprint anti-counterfeiting effect of this method is better.

[0157] This OCT fingerprint anti-counterfeiting method based on physiological structure consistency constructs a dual-branch model to identify the OCT image of the fingerprint to be detected, obtaining two intermediate features and two segmentation results. By calculating the similarity between the two intermediate features and the two segmentation results, it effectively distinguishes between genuine and counterfeit fingerprints. In the second stage of training, this OCT fingerprint anti-counterfeiting method based on physiological structure consistency forces the second intermediate feature output by the second segmentation network to align with the second intermediate feature output by the first segmentation network. This constraint ensures that the dual-branch model has highly similar feature representations to real samples, enabling it to not only effectively detect known forgery methods but also to achieve robust defense against unknown and novel forgery attacks by keenly capturing anomalies that deviate from the distribution of real physiological structures. The first and second loss functions of this method both adopt the dynamic weighted cross-entropy loss function, dynamically adjusting the weights according to the Dice coefficient on the validation set in each round, giving higher suggestion weights to target categories with poor segmentation performance, and forcing the dual-branch model to pay equal attention to the physiological structure of all categories.

[0158] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0159] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0160] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the 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 modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. An OCT fingerprint anti-counterfeiting method based on physiological structural consistency, characterized in that: The OCT fingerprint anti-counterfeiting method based on physiological structural consistency includes: OCT images of different fingerprints were collected, and ROI images were extracted to form a dataset. The dataset was divided into a training set and a validation set, and a portion of the ROI images in the training set and the validation set were labeled. A dual-branch model is constructed and trained in stages. The dual-branch model includes a first segmentation network and a second segmentation network with the same structure. Each segmentation network includes an extractor and a reconstructor connected in sequence. The extractor outputs intermediate features, and the reconstructor outputs segmentation results about the physiological structure of the skin. The segmentation results include segmented images of four categories: background, stratum corneum, sweat glands and active epidermal junction layer. During the training of the dual-branch model, in the first stage, the first segmentation network is trained under full supervision using a labeled training set to obtain the trained first segmentation network. In the second stage, the first step is to train the extractor in the second segmentation network using an unlabeled training set, and the second step is to train the extractor and reconstructor of the second segmentation network using a labeled training set. The first and second steps are executed alternately to obtain the trained second segmentation network. The OCT image of the fingerprint to be detected is acquired, and after the preset ROI image is extracted, it is input into the trained first segmentation network and second segmentation network respectively to obtain two first intermediate features and two first segmentation results. The anti-counterfeiting score is calculated, and the anti-counterfeiting score is used to determine whether the OCT image to be detected comes from a real finger, thereby realizing fingerprint anti-counterfeiting.

2. The OCT fingerprint anti-counterfeiting method based on physiological structural consistency as described in claim 1, characterized in that: The process of acquiring the ROI image includes: Obtain OCT scan data of the fingerprint region of the finger to obtain the fingerprint OCT image. Calculate the average pixel brightness of each row of the OCT image and select the row with the highest brightness. Line, using a width of The sliding window is for the brightest area. The rows are cropped according to a preset overlap rate, and the resulting image after each cropping is the [image type / description]. The ROI image.

3. The OCT fingerprint anti-counterfeiting method based on physiological structural consistency as described in claim 1, characterized in that: OCT images of different fingerprints were acquired, and ROI images were extracted. Each ROI image was then stretched and deformed, and finally adjusted using an interpolation algorithm. The dataset consists of all ROI images with the specified resolution and adjusted by the interpolation algorithm. The annotation process includes: annotating the pixels in the ROI image that belong to the background, stratum corneum, sweat glands, and active epidermal junction layer, respectively.

4. The OCT fingerprint anti-counterfeiting method based on physiological structure consistency as described in claim 1, characterized in that: The extractor includes a convolutional block, a first downsampling module, a second downsampling module, and a third downsampling module connected in sequence; The convolutional block and each downsampling module include a convolutional layer, a batch normalization layer and an activation layer connected in sequence; The reconstructor includes a first upsampling module, a second upsampling module, a third upsampling module, and an output prediction block connected in sequence; Each of the upsampling modules includes a transposed convolutional layer, a batch normalization layer, and an activation layer connected in sequence, and the output prediction block includes a convolutional layer, an activation layer, and a binarization process connected in sequence.

5. The OCT fingerprint anti-counterfeiting method based on physiological structural consistency as described in claim 1, characterized in that: The step of using a labeled training set to perform fully supervised training on the first segmentation network to obtain the trained first segmentation network includes: Before training, the number of pixels belonging to the background, stratum corneum, sweat glands and active epidermal connective layer in the labeled training set is counted. Initial weights are set for any category among the background, stratum corneum, sweat glands and active epidermal connective layer to obtain the initial weights of each category. The initial weights are used as the weights of each category in the first round of training. During a round of training: The first segmentation network is forward propagated using a labeled training set to obtain the first predicted segmentation image of each sample in the labeled training set corresponding to the four categories. Perform a first normalization on the weights of each category in the current round; The first loss function is calculated based on the first predicted segmented image of the four categories, the real segmented image, and the weights of the four categories after normalization in the current round. The parameters of the first segmentation network are then updated through backpropagation. The performance of the first segmentation network was evaluated using a validation set, and the first Dice coefficients for the four categories were calculated using the validation set during the evaluation process. Calculate the first suggested weight for each category based on the first Dice coefficient for each category, and update the first normalized weights for each category in the current round based on the first suggested weights for each category. The updated weights for each category will then be used as the weights for each category in the next round of training. Enter the next round of training, and continue the cycle until the first Dice coefficients of each category converge, thus obtaining the trained first segmentation network.

6. The OCT fingerprint anti-counterfeiting method based on physiological structure consistency as described in claim 5, characterized in that: When the first and second steps are executed alternately, each execution of the first and second steps in sequence is called a round. For each round: The first step is to input each sample in the unlabeled training set into the extractors of the second segmentation network and the trained first segmentation network respectively to obtain two second intermediate features; The difference between the two second intermediate features is calculated using the second loss function, and the parameters of the extractor in the second segmentation network are updated by backpropagation. The second step is to perform forward propagation of the labeled training set on the second segmentation network to obtain the second predicted segmentation image corresponding to the four categories for each sample in the labeled training set. The weights of each category in the current round are normalized a second time, while the weights of each category in the first round use the initial weights. The third loss function is calculated based on the second predicted segmentation image of four categories, the real segmentation image, and the weights of the four categories after second normalization in the current round, and the parameters of the second segmentation network are updated by backpropagation. The performance of the second segmentation network was evaluated using a validation set, and the second Dice coefficients for the four categories were calculated using the validation set during the evaluation process. The second suggested weights for each category are calculated based on the second Dice coefficients for each category. The second suggested weights for each category are then updated based on the second suggested weights for each category in the current round. The updated weights for each category are then used as the weights for each category in the next round. Enter the next round, and continue the loop until the second loss function and the second Dice coefficients of each category converge, resulting in the trained second segmentation network.

7. The OCT fingerprint anti-counterfeiting method based on physiological structural consistency as described in claim 1, characterized in that: When performing preset ROI image extraction on the OCT image of the fingerprint to be detected, a width of [missing information] is used. The sliding window is the brightest part of the OCT image of the fingerprint to be detected. Extracting the center region of the row yields a size of... The preset ROI images are used as inputs to the trained first and second segmentation networks, respectively.

8. The OCT fingerprint anti-counterfeiting method based on physiological structural consistency as described in claim 1, characterized in that: When calculating the anti-counterfeiting score, the cosine distance between two first intermediate features is first calculated, and then a third normalization is performed to obtain the intermediate feature similarity score. Then, the third Dice coefficients are calculated between the segmented images of the two stratum corneum, the segmented images of the two sweat glands, and the segmented images of the two active epidermal junction layers in the two first segmentation results. Three third Dice coefficients are obtained, and the three third Dice coefficients are normalized in the fourth way to obtain the segmentation similarity score. The anti-counterfeiting score is obtained by multiplying the intermediate feature similarity score and the segmentation similarity score.

9. The OCT fingerprint anti-counterfeiting method based on physiological structural consistency as described in claim 1, characterized in that: The method of determining whether the OCT image to be detected comes from a real finger based on the anti-counterfeiting score includes: The anti-counterfeiting score is compared with a preset value. If the score is less than the preset value, the OCT image to be detected is from a real finger; otherwise, the OCT image to be detected is not from a real finger.

10. The OCT fingerprint anti-counterfeiting method based on physiological structural consistency as described in claim 1, characterized in that: The first and third loss functions both use the dynamic weighted cross-entropy loss function, while the second loss function uses cosine distance loss and L1 distance loss.

Images