A palm pathological line recognition method

Abstract

The application discloses a palm pathological line recognition method, and the recognition steps comprise the following steps: acquiring a palm pathological line image, inputting the palm pathological line image into a visual Mamba main network, and extracting n-level features of the palm pathological line; the n-level features and n-1-level features are fused through pyramid pooling attention; and based on the fused features, the palm pathological line category is obtained through cross attention in combination with a learnable label embedding matrix. The application can effectively extract multi-scale features to recognize small-scale pathological line features in the palm image, enhance the association between semantic labels and image regions and context modeling, and improve the accuracy of multi-label palm pathological line recognition.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and more specifically to a method for recognizing pathological patterns on the palm, which is particularly suitable for small-scale pathological pattern feature detection and multi-label classification tasks in the context of traditional Chinese medicine palm pathological pattern differentiation. Background Technology

[0002] Palm pathological pattern recognition is a key technology in traditional Chinese medicine image analysis. It aims to provide auxiliary evidence for the diagnosis of systemic diseases by analyzing the morphological characteristics and distribution patterns of eight typical pathological patterns on the palm skin. Traditional methods often rely on human experience or rule-based feature extraction (such as grayscale statistics and texture filtering), resulting in strong subjectivity and poor generalization ability.

[0003] While deep learning-based methods have made progress in recent years, they still face significant challenges in multi-label palmar pathological print recognition tasks. For example:

[0004] Existing convolutional neural networks (CNNs) extract features by stacking convolutional layers with fixed receptive fields, making it difficult to simultaneously capture both local fine lines and global distribution patterns. In particular, they are insufficient in modeling the geometric characteristics of pathological lines, such as directionality and continuity, leading to a high false negative rate for small-scale pathological lines. While the Visual Transformer (ViT) can model long-range dependencies, its self-attention computational complexity increases quadratically with image resolution, making it difficult to meet the efficient processing requirements of high-resolution palm images (e.g., 448×448). Multilayer perceptron models model global relationships through fully connected layers, but lack the ability to model the spatial association between local details and labels, further limiting the accuracy of fine-grained recognition.

[0005] Furthermore, in palm pathological pattern recognition, multi-label classification requires explicit association between labels and image regions. However, existing methods often use global pooling or fully connected layers to directly map features, ignoring the spatial dependencies and semantic hierarchies between labels. For example, the co-occurrence region association between "cross" patterns and "rice" patterns is not effectively modeled, leading to deviations in prediction results.

[0006] Furthermore, palm pathological print recognition faces challenges related to data label imbalance and training optimization: the long-tailed distribution leads to a significant scarcity of rare pathological samples, while the traditional cross-entropy loss exacerbates the tendency of overfitting the head label; existing weighted losses (such as ASL asymmetric loss) alleviate class imbalance by adjusting the weights of positive and negative samples, but do not consider the guiding role of label co-occurrence relationships in gradient allocation, and a single loss function is difficult to balance accuracy and stability; conventional data augmentation operations (such as random cropping and rotation transformation) easily destroy the continuity of pathological prints, and cross-domain data differences further weaken the model's generalization ability.

[0007] Representative methods such as the Query2Label framework rely on single-scale feature decoding and lack the integration of multi-level information, resulting in insufficient accuracy in recognizing small-scale pathological patterns. The Swin-Transformer is too computationally expensive to adapt to high-resolution palm images and lacks optimization for scanning strategies for anisotropic patterns.

[0008] The above-mentioned scheme has defects in feature fusion efficiency, label association, and loss optimization, which ultimately restricts the improvement of the accuracy of multi-label pathological pattern recognition. Summary of the Invention

[0009] In view of this, in order to at least partially solve the above-mentioned technical problems, the present invention provides a multi-label palm pathological print recognition method based on multi-stage pyramid pooling attention feature fusion.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A method for identifying pathological patterns on the palm, comprising the following steps:

[0012] Acquire images of pathological patterns on the palm, input them into a visual Mamba backbone network, and extract n-level features of the pathological patterns on the palm.

[0013] Features at level n and level n-1 are fused using pyramid pooling attention;

[0014] Based on fusion features and combined with a learnable label embedding matrix, the palm pathological pattern category is obtained through cross attention.

[0015] Furthermore, the visual Mamba backbone network includes block embedding layers and multi-level feature extraction layers; each feature extraction layer includes multiple stacked Inception-SSM modules.

[0016] Furthermore, the Inception-SSM module includes a channel splitting layer and a channel splicing layer, with a first branch and a second branch connected in parallel between the channel splitting layer and the channel splicing layer;

[0017] The first branch includes a preprocessing layer and a multi-core Inception module;

[0018] The second branch includes a fully connected projection layer, and connected to it a depth-separable convolutional layer, a first SiLU (Sigmoid-Weighted Linear Unit) activation layer, an SS2D scanning layer, and a channel attention layer; at the same time, the fully connected projection layer is connected to the fusion branch layer together with the channel attention layer through the second SiLU activation layer.

[0019] Furthermore, the features at level n and level (n-1) are fused using pyramid pooling attention, including:

[0020] Align the nth-level feature space with the (n-1)th-level feature space and perform initial fusion;

[0021] The initial fused features are adjusted using adaptive channel attention;

[0022] The adjusted features are subjected to multi-scale pooling, and the generated multi-scale pyramid feature maps are positionally encoded and concatenated.

[0023] The concatenated features are used as keys and values, and the nth-level features are used as query vectors. Feature fusion is achieved through a multi-head attention mechanism.

[0024] As a preferred option, before spatially aligning the nth-level features with the (n-1)th-level features, they are pre-processed through 1×1 convolutional layers to map the number of feature channels and unify the dimensions.

[0025] Furthermore, the spatial alignment of the nth-level feature with the (n-1)th-level feature includes:

[0026] The nth level feature is passed sequentially through a bilinear interpolation layer, a depthwise separable convolutional layer, a normalized activation layer, and a convolutional layer.

[0027] Furthermore, the initial fused features are adjusted through adaptive channel attention, including:

[0028] The initial fused features are passed sequentially through a global pooling layer, a 1D convolutional layer, and a function activation layer to obtain attention weights;

[0029] Attention weights are used to adjust the initial fusion features.

[0030] As a preferred approach, the initial fusion features are first input into a depth-separable convolutional layer for processing before adjustment, in order to reduce the amount of computation / parameters.

[0031] Furthermore, based on the fused features and combined with the learnable label embedding matrix, palmar pathological pattern categories are obtained through cross-attention, including:

[0032] The fused features are used as keys and values, and the learnable label embedding matrix is ​​used as a query. The palm pathological pattern category feature vector is output through cross attention.

[0033] The categorical feature vectors are normalized using the Sigmoid function to generate multi-label prediction results.

[0034] Furthermore, the recognition process is optimized in two stages;

[0035] The first stage uses an unbalanced weighted loss function:

[0036]

[0037] In the formula, K represents the total number of label categories, i represents the label category index in the first stage, α is the loss balance adjustment parameter, γ represents the weight of the adjustment parameter, and y i y represents the true label of sample i, and y represents a positive sample. i =1, negative sample y i =0, p i This represents the probability that the model predicts sample i belongs to the positive class.

[0038] The second stage uses an asymmetric loss function:

[0039]

[0040] In the formula, K represents the total number of tag categories, k represents the tag category index in the second stage, and y k It is a binary label indicating whether the input palmar pathological print image contains the label k, p k This indicates the predicted probability that the image contains class k. The total loss is calculated by averaging the losses of all samples in the training set; in the above formula, γ + and γ - There are two focal parameters, γ + It is a modulating factor for positive samples, responsible for adjusting the model's weights on easily classified positive samples. When the predicted probability of a positive sample approaches 1, the loss function value decreases accordingly, while γ... - The negative sample adjustment factor is used to increase the penalty for negative samples that are difficult to classify by the model.

[0041] By setting γ + <γ - This allows for better control over the contribution of harder samples during training.

[0042] Furthermore, the optimized total loss function is as follows:

[0043]

[0044] In the formula, This represents the loss function in the first stage. Let β represent the loss function for the second stage, and let β represent the weight parameter.

[0045] Furthermore, a multi-label dataset of palmar pathological prints was constructed, key palm regions were extracted and data augmentation was performed, and the enhanced data was used to optimize the recognition process.

[0046] Data augmentation includes:

[0047] The mixing ratio was determined based on the Beta distribution and the uniform distribution.

[0048] Based on the mixing ratio, perform linear interpolation on the images of the current batch and the randomly shuffled batch, or perform local replacement within a random region;

[0049] Assign smooth soft labels to images that have undergone linear interpolation or local replacement.

[0050] This invention discloses a method for identifying pathological patterns on the palm, which, compared with existing technologies,

[0051] 1. The visual Mamba backbone network can effectively extract multi-scale local details and global spatial features. By adopting a multi-path selection scanning strategy, the model's ability to model the anisotropic features of pathological patterns is effectively improved.

[0052] 2. The multi-stage pyramid pooling attention feature fusion mechanism of this application enhances the semantic association and contextual modeling ability between features, which helps to identify small-scale pathological texture features;

[0053] 3. This invention mitigates the negative impact of data label imbalance by employing a two-stage dynamic optimization strategy using an unbalanced weight loss function and an asymmetric loss function, thereby improving the model's ability to identify tail categories while maintaining training stability.

[0054] In summary, the innovative techniques proposed in this invention, combined with other technologies, significantly improve the performance of multi-label palmar pathological print recognition, demonstrating substantial improvements in both classification accuracy and computational efficiency. On the palmar pathological print dataset, the mean accuracy (mAP) is improved by 1%, and inference efficiency is increased by 25%. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0056] Figure 1 The flowchart of the palm pathological print recognition method provided by the present invention;

[0057] Figure 2 Visual Mamba backbone network architecture diagram;

[0058] Figure 3 Here is a structural diagram of the Inception-SSM module;

[0059] Figure 4 This is a schematic diagram of a multi-stage pyramid pooling attention fusion module;

[0060] Figure 5A flowchart illustrating the computation of cross-attention for embedding the encoder into the label;

[0061] Figure 6 This is a schematic diagram of the two-stage loss function optimization training. Detailed Implementation

[0062] 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.

[0063] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0064] This invention discloses a method for recognizing pathological patterns on the palm, which can not only effectively extract multi-scale features to identify small-scale pathological pattern features in palm images and enhance the association and contextual modeling between semantic labels and image regions, but also alleviate the class learning bias caused by label imbalance through two-stage loss optimization, thereby improving the accuracy of the model in recognizing pathological patterns on multi-label palms.

[0065] like Figure 1 The steps include:

[0066] Acquire images of pathological patterns on the palm, input them into a visual Mamba backbone network, and extract n-level features of the pathological patterns on the palm.

[0067] Features at level n and level n-1 are fused using pyramid pooling attention;

[0068] Based on fusion features and combined with a learnable label embedding matrix, the palm pathological pattern category is obtained through cross attention.

[0069] In one embodiment:

[0070] The visual Mamba backbone network consists of patch embedding layers and multi-level feature extraction layers;

[0071] Block embedding is used to convert a two-dimensional image into a series of vectors so that the network can process image data in sequence.

[0072] Each feature extraction layer consists of multiple stacked Inception-SSM modules.

[0073] This embodiment uses four-level feature extraction as an example for illustration. In this case, the visual Mamba backbone network structure is as follows: Figure 2 As shown, it consists of a Patch Embedding layer and four deep feature extraction stages; each deep feature extraction stage is connected by convolutional downsampling to reduce the spatial dimension of the feature map, thereby reducing the computational cost and the number of parameters.

[0074] Furthermore, each Inception-SSM module includes a channel splitting layer and a channel splicing layer, with the structure referring to... Figure 3 The first branch and the second branch are connected in parallel between the channel segmentation layer and the channel splicing layer;

[0075] The first branch includes a preprocessing layer and a multi-core Inception module;

[0076] The multi-core Inception module uses residual connections to preserve the original features, captures local features through 3×3 depthwise separable convolutional layers, captures horizontal features through 1×11 horizontal strip convolutions, and captures vertical features through 11×1 vertical strip convolutions. After concatenating the multi-branch feature outputs along the channel dimension, feature adjustment is completed through 1×1 convolution.

[0077] The second branch includes a fully connected projection layer, and connected to it a depth-separable convolutional layer, a first SiLU activation layer, an SS2D scanning layer, and a channel attention layer; at the same time, the fully connected projection layer is connected to the fusion branch layer together with the channel attention layer through the second SiLU activation layer.

[0078] The SS2D scanning layer is used to unfold two-dimensional features into one-dimensional sequences along the four directions: horizontal, vertical, and diagonal (sequence dimensions are set to 48, 96, 192, and 384 in stages one through four, respectively). Each direction forms an independent scanning path, and is processed in parallel through selective state space scanning (S6) to ensure that each pixel integrates multi-directional contextual information through compressed hidden states (hidden state dimension is 16).

[0079] In stage S6 of the SS2D scanning layer, the input sequence is divided into two parts: one part, x, is used for core scan computation, and the other part, z, is used for subsequent gating mechanisms. The core of S6 lies in its state transition matrices A, B, and C, as well as the discretization step size Δ, which are dynamically generated based on the input x. The input features x dynamically adjust the state transition matrices and the hidden state update mechanism to achieve selective feature updates. Furthermore, to process time series data on a digital computer, the sequence needs to be discretized. The formula for calculating the hidden state is as follows:

[0080]

[0081] Mamba transforms the dynamic state transition matrices A and B into a zero-order hold (ZOH) method. and In the formula, It was determined that from the previous state h t-1 The amount of information inherited (selective forgetting), while The current input x is set. t Contribution to state updates (selective writes). Furthermore, via Δ t It controls the degree of discretization of A. A large Δ t This means that the system responds to the current input x t More sensitive, and the previous state h t-1 The effect may decay at a faster rate (depending on the eigenvalues ​​of matrix A). Small Δ t The focus is more on maintaining the previous state h. t-1 .

[0082] y at each time step scan_t The output is determined by the current hidden state h. t It is determined by the dynamic output matrix C, and the formula is as follows:

[0083] y scan_t =CH t

[0084] Although the above formulas appear sequential, Mamba actually employs a parallel scan algorithm to model long-range dependencies, ensuring that the computational task can be completed in O(n) time complexity. Subsequently, the computational results output by the S6 module are integrated into y. scan The gating value z is then normalized and fused with the previously separated gating value z (e.g., through element-wise multiplication and an activation function such as SiLU), thereby significantly improving the model's expressiveness and selectivity. The gating calculation method is as follows:

[0085] output = y scan t ×SiLU(z)

[0086] In a specific embodiment, when given an input palm image of 448×448×3, it is first downsampled to 112×112×96 by the PatchEmbedding layer after passing through the visual Mamba backbone neural network. Then, the network extracts features through parallel modules at different stages, downsampling to 56×56×192 and 28×28×384 step by step. Finally, the output image spatial feature map size is 14×14×768.

[0087] In one embodiment:

[0088] To enhance the semantic association and context awareness of multi-scale features and ensure that small-scale pathological patterns receive sufficient attention, this application fuses the nth and (n-1)th level features using pyramid pooling attention. The steps include:

[0089] 1. Align the nth-level feature with the (n-1)th-level feature spatially and perform initial fusion; wherein, spatial alignment is to pass the nth-level feature sequentially through a bilinear interpolation layer, a depthwise separable convolutional layer, a normalized activation layer, and a convolutional layer, and then multiply it element-wise with the (n-1)th-level feature;

[0090] As a preferred implementation, before spatial alignment of the nth-level features and the (n-1)th-level features, they are pre-passed through a 1×1 convolutional layer to reduce the number of feature channels.

[0091] 2. Adjust the initial fused features through adaptive channel attention; that is, the initial fused features pass through the global pooling layer, the 1D convolutional layer and the function activation layer in sequence to obtain attention weights; and then use the attention weights to adjust the initial fused features by weighting.

[0092] As a preferred implementation, the preliminary fusion features are first input into a depth-separable convolutional layer for processing before adjustment, in order to reduce the amount of computation / parameters.

[0093] 3. Perform multi-scale pooling on the adjusted features, and encode and concatenate the generated multi-scale pyramid feature maps.

[0094] 4. Use the concatenated features as keys and values, and the nth-level features as query vectors, to achieve feature fusion through multi-head attention.

[0095] In one exemplary embodiment, such as Figure 4 First, the 14×14×768 feature map output from the fourth stage is dynamically upsampled and spatially aligned with the 28×28×768 feature map from the third stage, and then initially fused.

[0096] The fused third-stage features are then weighted using a 3×3 depthwise separable convolution and an adaptive channel attention mechanism to achieve feature adjustment; the attention weights are calculated using the following formula:

[0097]

[0098] in, Let W represent the input feature map, GAP(·) represent the global average pooling operation, and W represent the input feature map. k Let σ(·) represent a one-dimensional convolution with kernel size k, and let σ(·) represent the Sigmoid activation function.

[0099] Finally, using the pyramid pooling attention module, the fourth-stage features are used as query vectors. The third-stage features, adjusted by the channel attention mechanism, are pooled at four scales: 2×2, 3×3, 5×5, and 7×7. The generated multi-scale pyramid feature maps are then encoded with relative positions and concatenated to form keys and values. Finally, deep fusion is achieved through a multi-head self-attention mechanism.

[0100] In one embodiment:

[0101] To improve the accuracy of multi-label classification, based on fused features and combined with a learnable label embedding matrix, the palm pathological pattern category is obtained through cross-attention. The steps include:

[0102] The fused features are used as keys and values, and the learnable label embedding matrix is ​​used as a query. The palm pathological pattern category feature vector is output through cross attention.

[0103] The categorical feature vectors are normalized using the Sigmoid function to generate multi-label prediction results.

[0104] In this embodiment, the process of identifying palmistry patterns refers to... Figure 5 First, a learnable label embedding matrix is ​​defined. (K is the number of label categories, d = 512), and the embedding vector of each label category is used as the query Q, and the fused multi-scale feature F fusion As K and V, the category response is calculated using the cross-attention module:

[0105]

[0106] Then, the output feature vector is normalized by the Sigmoid function to generate multi-label prediction results.

[0107] In one embodiment:

[0108] Before recognizing palm prints, the above recognition process is trained and optimized.

[0109] In this embodiment, to improve the diversity of training samples, relevant data were first collected for eight common types of pathological patterns on the palm (grid pattern, cross pattern, star pattern, triangle pattern, island pattern, chain pattern, star pattern, and grid pattern) to construct a multi-label dataset of palm pathological patterns. Then, the palm region was accurately located and extracted using a pre-trained YOLOv8 palm target detection model, and cropped and standardized to a uniform size of 448×448 pixels.

[0110] To alleviate the problems of data scarcity and class imbalance, this application implements a hybrid enhancement strategy at the batch dimension, specifically as follows:

[0111] A random mixing ratio λ is determined, and linear interpolation or local replacement is performed on the images of the current batch and the randomly shuffled batches. Then, smooth soft labels are generated based on the mixing ratio λ.

[0112] y mix =λy1+(1-λ)y2

[0113] Where λ is the random mixing coefficient.

[0114] The initial distribution of the λ parameter is determined by a hybrid strategy: 50% probability of sampling from the Beta(0.8, 0.8) distribution (linear interpolation enhancement, with values ​​concentrated in the 0.3-0.7 interval), and 50% probability of sampling from the uniform distribution (local replacement enhancement, with equal probability across the entire interval).

[0115] Furthermore, the recognition process is optimized in two stages to mitigate the impact of label imbalance on model training, such as... Figure 6 As shown;

[0116] In the first stage, the output features of the backbone network are processed by the Sigmoid function to generate the first classification probability, and the first loss value is calculated using the balanced weight loss function.

[0117]

[0118] In the formula, K represents the total number of label categories, i represents the label category index in the first stage, α is the loss balance adjustment parameter, γ represents the weight of the adjustment parameter, and y i y represents the true label of sample i, and y represents a positive sample. i =1, negative sample y i =0, p i This represents the probability that the model predicts sample i belongs to the positive class.

[0119] In the second stage, the output features of the cross-attention module are processed by the Sigmoid function to generate a second classification probability, and the second loss value is calculated using an asymmetric loss function.

[0120]

[0121] In the formula, K represents the total number of tag categories, k represents the tag category index in the second stage, and y k It is a binary label indicating whether the input palmar pathological print image contains the label k, p k This represents the predicted probability that the image contains class k. The total loss is calculated by averaging the losses of all samples in the training set. In the above formula, γ + and γ - There are two focal parameters, γ +It is a modulating factor for positive samples, responsible for adjusting the model's weights on easily classified positive samples. When the predicted probability of a positive sample approaches 1, the loss function value decreases accordingly, while γ... - The negative sample adjustment factor is used to increase the penalty for negative samples that are difficult to classify by the model.

[0122] By setting γ + <γ - This allows for better control over the contribution of harder samples during training.

[0123] Finally, the weighted total loss is backpropagated using gradient descent to optimize the model parameters. The optimized total loss function is:

[0124]

[0125] In the formula, This represents the loss function in the first stage. Let β represent the loss function for the second stage, and let β represent the weight parameter.

[0126] This application effectively reduces the negative impact of label imbalance on model optimization by dynamically weighting a loss function based on unbalanced weights and an asymmetric loss function. Initially, the unbalanced weights loss function is used as the primary method, gradually transitioning to an asymmetric loss function later. The smooth switching of the loss function is achieved by dynamically scheduling the parameter β, further improving the model's generalization ability and stability.

[0127] In an embodiment of the present invention, the parameters for model training are configured as follows: the batch size is set to 16, and the initial learning rate is 5 × 10⁻⁶. -5 The AdamW optimization algorithm and cosine learning rate scheduler were selected. In the initial training phase, specifically the first 10 epochs, a cosine annealing strategy was used for warm-up to gradually adjust the learning rate, thereby improving the model's convergence stability. On an RTX 3090 GPU, the total number of training epochs was set between 100 and 200, and an early stopping strategy was implemented.

[0128] The identification method of the present invention can be executed by a computer program to achieve disease identification based on a trained dedicated medical image dataset, effectively extracting multi-scale pathological features and achieving a high classification accuracy.

[0129] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0130] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying pathological patterns on the palm, characterized in that, include: Acquire images of pathological patterns on the palm, input them into a visual Mamba backbone network, and extract n-level features of the pathological patterns on the palm. The visual Mamba backbone network includes a block embedding layer and a multi-level feature extraction layer; each feature extraction layer includes multiple stacked Inception-SSM modules; the Inception-SSM module includes a channel segmentation layer and a channel splicing layer, with a first branch and a second branch connected in parallel between the channel segmentation layer and the channel splicing layer; The first branch includes a preprocessing layer and a multi-core Inception module; The second branch includes a fully connected projection layer, and connected to it a depth-separable convolutional layer, a first SiLU activation layer, an SS2D scanning layer, and a channel attention layer; at the same time, the fully connected projection layer is connected to the fusion branch layer together with the channel attention layer through the second SiLU activation layer. Features at level n and level (n-1) are fused using pyramid pooling attention; including: Align the nth-level feature with the (n-1)th-level feature space and perform initial fusion; the alignment of the nth-level feature with the (n-1)th-level feature space includes: passing the nth-level feature sequentially through a bilinear interpolation layer, a depthwise separable convolutional layer, a normalized activation layer, and a convolutional layer; The initial fused features are adjusted using adaptive channel attention; this includes: passing the initial fused features sequentially through a global pooling layer, a 1D convolutional layer, and a function activation layer to obtain attention weights; and using these attention weights to perform weighted adjustments on the initial fused features. The adjusted features are subjected to multi-scale pooling, and the generated multi-scale pyramid feature maps are then positionally encoded and concatenated. The concatenated features are used as keys and values, and the nth-level features are used as query vectors. Feature fusion is achieved through multi-head attention. Based on fusion features and combined with a learnable label embedding matrix, the palm pathological pattern category is obtained through cross attention.

2. The palm pathological print recognition method according to claim 1, characterized in that, Based on fused features and combined with a learnable label embedding matrix, palm pathological pattern categories are obtained through cross-attention, including: The fused features are used as keys and values, and the learnable label embedding matrix is ​​used as a query. The palm pathological pattern category feature vector is output through cross attention. The categorical feature vectors are normalized using the Sigmoid function to generate multi-label prediction results.

3. The method for identifying pathological palm prints according to claim 1, characterized in that, The recognition process is optimized in two stages; The first stage uses an unbalanced weighted loss function: ； In the formula, K represents the total number of tag categories, and i represents the tag category index in the first stage. For loss balance adjustment parameters, This indicates the adjustment of parameter weights. This represents the true label of sample i, when it is a positive sample. negative samples , This represents the probability that the model predicts sample i belongs to the positive class; The second stage uses an asymmetric loss function: ； In the formula, K represents the total number of tag categories, and k represents the tag category index in the second stage. It is a binary label indicating whether the input palmar pathological print image contains label k. This indicates the predicted probability that the current image contains category k. and These are two focus parameters. This is a moderating factor for positive samples, responsible for adjusting the model's weights on easily classified positive samples. As the predicted probability of a positive sample approaches 1, the loss function value decreases accordingly. It is a negative sample adjustment factor used to increase the penalty for difficult-to-classify negative samples by the model.

4. The palm pathological print recognition method according to claim 3, characterized in that, The optimized total loss function is: ； In the formula, This represents the loss function in the first stage. Let β represent the loss function for the second stage, and let β represent the weight parameter.

5. The method for identifying pathological patterns on the palm according to claim 1, characterized in that, Construct a multi-label dataset of palmar pathological prints, extract key palm regions, and perform data augmentation. And to optimize the recognition process using the enhanced data; Data augmentation includes: The mixing ratio was determined based on the Beta distribution and the uniform distribution. Based on the mixing ratio, perform linear interpolation on the images of the current batch and the randomly shuffled batch, or perform local replacement within a random region; Assign smooth soft labels to images that have undergone linear interpolation or local replacement.

Images