Adaptive noise suppression cooperative coupling feature enhanced palm vein fold point detection method

Abstract

The application provides a palm vein fold point detection method with adaptive noise suppression and coupled feature enhancement, acquires a palm vein image to be processed; inputs the palm vein image into a trained fold point detection network to obtain the number and position information of the palm vein fold points, and the fold point detection network comprises a cascaded shallow feature extraction module, an adaptive noise suppression module, a coupled feature enhancement module and a decoder. The method realizes the suppression of the background noise of the palm vein image through the adaptive noise suppression module, then enhances the spatial dependence relationship between channels through the coupled feature enhancement module to realize the strengthening of the position information, simultaneously increases the kernel size from bottom to top through the pyramid structure and reduces the kernel depth through grouped convolution, realizes the capture of multi-scale information, and thus realizes the effective and accurate detection of the number and position of the fold points in the palm vein image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a palm vein fold point detection method with adaptive noise suppression and coupled feature enhancement. BACKGROUND

[0002] With the continuous development of science and technology, biometric recognition technology is increasingly widely used in various fields of society, especially in the fields of identity authentication and security monitoring. As a unique and difficult-to-fake biometric feature, palm vein has gradually become an emerging identity verification technology due to its high individual difference, and is applied to many scenarios such as financial payment, smart device unlocking, public security, etc. In the financial field, the introduction of palm vein technology can improve the security of transactions and prevent identity theft. In smart home and health monitoring, palm vein can also provide a convenient and secure individual identification method to enhance the intelligence and personalization of devices. In addition, the analysis of palm vein details can also be used for health management to help early detection of signs of certain diseases. Therefore, efficient palm vein recognition and fold point analysis can provide a safer and smarter living environment for society, and promote the innovation and popularization of biometric recognition technology. The existing palm vein detection methods have problems such as insufficient background noise suppression, loss of target position information, and difficulty in extracting multi-scale features of images, which leads to low accuracy of palm vein fold point detection. SUMMARY

[0003] In view of the defects of the prior art, the present application provides a palm vein fold point detection method with adaptive noise suppression and coupled feature enhancement to solve the technical problem of inaccurate palm vein fold point detection.

[0004] To achieve the above object and other related objects, the present application provides a palm vein fold point detection method with adaptive noise suppression and coupled feature enhancement.

[0005] In an embodiment of the present application, the method comprises: acquiring a palm vein image to be processed; inputting the palm vein image into a trained fold point detection network to obtain the number and position information of palm vein fold points, wherein the fold point detection network comprises a cascade of: a shallow feature extraction module for preprocessing the palm vein image; an adaptive noise suppression module for suppressing background noise to improve the signal-to-noise ratio of the image; a coupled feature enhancement module for enhancing the position information between each channel to obtain multi-scale information; and a decoder for processing the enhanced features to obtain the number and position information of the palm vein fold points.

[0006] In an embodiment of the present application, the adaptive noise suppression module processes the shallow features output by the shallow feature extraction module according to the following steps: divides the shallow features X into G sub-features Xi , i is in {1, 2, …, G}; the sub-feature X i is processed by using a GAP algorithm to obtain a first feature X 1i ; the first feature X 1i is normalized by using a constant epsilon to obtain a second feature X 2i ; the second feature X 2i is transformed by using learnable parameters gamma and beta to obtain a third feature X 3i ; the third feature X i is processed by using the sub-feature X 3i and a Sigmoid function to obtain a fourth feature X 4i ; and the fourth feature X 4i is spliced to obtain an output feature Y of the adaptive noise suppression unit.

[0007] In an embodiment of the present application, the relationship between each feature is expressed by the following formula:

[0008] X 1i =X i ·GAP(X i );

[0009] ;

[0010] X 3i =gamma X 2i +beta;

[0011] X 4i =X i ·Sigmoid(X 3i );

[0012] Y=X 41 +X 42 +…+X 4G .

[0013] In an embodiment of the present application, the coupling feature enhancement unit includes a learnable graph neural unit and a multi-scale information fusion unit, the learnable graph neural unit enhances the spatial dependency relationship between each channel by using a graph structure, and the multi-scale information fusion unit captures multi-scale information by using a double-oriented pyramid operation.

[0014] In an embodiment of the present application, the expression of the learnable graph neural unit is as follows:

[0015] Z=Ycdot ReLU(Conv1d(Acdot GAP(Y)));A=A0x A1+A2;

[0016] A0 = Softmax(Conv1d(GAP(Y))); where Z is the output feature of the learnable graph neural unit, Y is the output feature of the adaptive noise suppression unit, GAP is the GAP algorithm, Conv1d is a one-dimensional convolutional layer, ReLU is the activation function, Softmax is the normalization, and A0, A1 and A2 are learnable adjacency matrices.

[0017] In one embodiment of the present invention, the multi-scale information fusion unit includes a parallel first branch and a second branch. The first branch consists of a cascaded first convolutional layer, multiple parallel group convolutions, and a second convolutional layer. The second branch consists of a cascaded adaptive average pooling layer, a first convolutional layer, multiple parallel group convolutions, a second convolutional layer, and a first upsampling layer. The output features of the first branch and the second branch are connected together and then sequentially passed through a third convolutional layer and a second upsampling layer to obtain the output features of the multi-scale information fusion unit.

[0018] In one embodiment of the present invention, the plurality of parallel grouped convolutions include four grouped convolutions: the first grouped convolution has a kernel size of 9×9 and a group size of 16; the second grouped convolution has a kernel size of 7×7 and a group size of 8; the third grouped convolution has a kernel size of 5×5 and a group size of 4; and the fourth grouped convolution has a kernel size of 3×3 and a group size of 1.

[0019] In one embodiment of the present invention, the loss function during training of the inflection point detection network is as follows: K Overall =K Bayesian +λK Count Among them, K Overall K is the total loss function. Bayesian For Bayesian uncertainty loss, K Count λ is the counting error loss, and λ is a hyperparameter used to balance the weights of the two losses.

[0020] The beneficial effects of this invention are as follows: This invention proposes a palm vein inflection point detection method based on adaptive noise suppression and synergistic coupling feature enhancement. This method uses an adaptive noise suppression module to suppress background noise in palm vein images, and then uses a coupling feature enhancement module to enhance the spatial dependency between channels to strengthen the positional information. At the same time, it uses a pyramid structure to increase the kernel size from bottom to top and grouped convolution to reduce the kernel depth, thereby capturing multi-scale information and achieving effective and accurate detection of the number and location of inflections in palm vein images. Attached Figure Description

[0021] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an architecture diagram of a kinetic detection network provided in an embodiment of the present invention;

[0023] Figure 2 This is a comparison image of palm veins before and after processing provided in an embodiment of the present invention;

[0024] Figure 3 This is an architecture diagram of an adaptive noise suppression module provided in an embodiment of the present invention;

[0025] Figure 4 This is an architectural diagram of a learnable graph neural unit provided in an embodiment of the present invention;

[0026] Figure 5 This is an architecture diagram of a multi-scale information fusion unit provided in an embodiment of the present invention. Detailed Implementation

[0027] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other. In addition to the specific methods, equipment, and materials used in the embodiments, based on the knowledge of the prior art and the description of the present invention by those skilled in the art, any prior art methods, equipment, and materials similar to or equivalent to those described in the embodiments of the present invention can be used to implement the present invention.

[0028] It should be understood that the terminology used in the embodiments of this invention is for describing specific particular implementations and not for limiting the scope of protection of this invention. Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art.

[0029] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In some embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0030] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functions, and operations that may be implemented in the methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0031] Please see Figure 1 , Figure 1 An adaptive noise suppression and collaborative coupling feature enhancement method for palm vein breakpoint detection, provided in one embodiment of the present invention, includes two steps: First, acquiring a palm vein image to be processed, such as... Figure 2 As shown on the left; next, the palm vein image is input into the trained inflection point detection network to obtain the number and location information of palm vein inflections. The inflection point detection network includes a cascaded shallow feature extraction module, an adaptive noise suppression module, a coupled feature enhancement module, and a decoder. The shallow feature extraction module is used to preprocess the palm vein image; the adaptive noise suppression module is used to suppress background noise to improve the signal-to-noise ratio of the image; the coupled feature enhancement module is used to enhance the positional information between channels to obtain multi-scale information; the decoder is used to process the enhanced features to obtain the number and location information of palm vein inflections. The processing result is shown below. Figure 2 As shown on the right.

[0032] This method suppresses background noise in palm vein images through an adaptive noise suppression module, enhances spatial dependencies between channels using a coupling feature enhancement module, and strengthens positional information by using a pyramid structure to increase kernel size from bottom to top and grouped convolution to reduce kernel depth, thereby capturing multi-scale information and achieving effective and accurate detection of the number and location of inflections in palm vein images.

[0033] Please see Figure 3 In a specific embodiment of the present invention, the adaptive noise suppression module processes the shallow features output by the shallow feature extraction module according to the following steps (1) to (6). Figure 3 F in att That is, the corresponding steps.

[0034] (1) Divide the shallow feature X into G sub-features X i , i∈{1,2,…,G}. In this step, the shallow features X∈R C ×H×W The feature space is then decomposed into G sub-feature modules using a higher-order decomposition to achieve more refined encoding and representation. This operation aims to extract richer local information through hierarchical segmentation of the original features, while enhancing the model's adaptability and recognition ability to diverse patterns, resulting in X. i ∈R C / G×H×W .

[0035] (2) Use the GAP algorithm to analyze sub-features X i The process is performed to obtain the first feature X. 1i GAP (Graph Attention Pooling) is an attention-based graph pooling algorithm used in graph neural networks (GNNs) to efficiently compress graph structures while retaining key information. Its core idea is to select the most important nodes or substructures in the graph using attention scores, thereby generating a more compact graph representation. In a specific embodiment of this invention, X... 1i =X i ·GAP(X i That is, using the GAP algorithm to analyze sub-features X i Processing yields GAP(X) i Then utilize sub-feature X i With GAP(X) i Multiplying them together yields the first characteristic X. 1i .

[0036] (3) Using the constant ε to represent the first characteristic X 1i Normalization is performed to obtain the second feature X. 2i In a specific embodiment of the present invention, the first feature X is first calculated. 1i The average value μ i and variance σ i Then, introducing a constant ε can avoid the second characteristic X. 2i The denominator is 0, and its specific formula is as follows:

[0037] .

[0038] (4) Using learnable parameters γ and β to model the second feature X 2i Perform the transformation to obtain the third feature X. 3i The purpose of this step is to ensure that unit transformations can be performed. In a specific embodiment of the invention, X 3i =γX 2i +β.

[0039] (5) Utilizing sub-feature X i and the Sigmoid function for the third feature X 3i After processing, the fourth feature X is obtained. 4i The activation function generates attention factors, and the refined features are generated by scaling the original sub-features. After optimizing each set of sub-features, they are integrated into the final optimized features to ensure that the feature depth matches the input. In a specific embodiment of the present invention, X 4i =X i ·Sigmoid(X 3i ).

[0040] (6) Splicing the fourth feature X 4i This yields the output feature Y of the adaptive noise suppression unit. It can be understood that for each sub-feature X... i After the above steps (2) to (5), a corresponding X can be obtained. 4i In the end, you will get G X's. 4i Finally, we only need to concatenate these fourth features along the original channel dimensions to obtain the overall feature Y, which can be expressed by the formula: Y=X 41 +X 42 +…+X 4G .

[0041] Through the above steps (1) to (6), the background noise of the image can be suppressed to a great extent, which is beneficial to the accurate extraction of features.

[0042] Please see Figure 1 In a specific embodiment of the present invention, the coupling feature enhancement unit includes a learnable graph neural unit and a multi-scale information fusion unit. The learnable graph neural unit uses graph structure to enhance the spatial dependency between each channel. The multi-scale information fusion unit uses pyramid structure to increase kernel size from bottom to top and reduces kernel depth (connectivity) through grouped convolution, that is, it captures multi-scale information through dual-guided pyramid operation.

[0043] Please see Figure 4In a specific embodiment of the present invention, the expression of the learnable graph neural unit is as follows: Z = Y • ReLU(Conv1d(A • GAP(Y))); A = A0 × A1 + A2; A0 = Softmax(Conv1d(GAP(Y))); where Z is the output feature of the learnable graph neural unit, Y is the output feature of the adaptive noise suppression unit, GAP is the GAP algorithm, Conv1d is a one-dimensional convolutional layer, ReLU is the activation function, Softmax is normalization, and A0, A1, and A2 are learnable adjacency matrices. In this embodiment, three adjacency matrices A0, A1, and A2 are used to represent the connections between vertices. A0 is a diagonal matrix based on self-attention, used to suppress useless features, and is generated by a one-dimensional convolutional layer Conv1d and a Softmax function; A1 contains the vertices themselves and needs to be normalized; A2 can customize unique dependencies for different feature vertices, and needs to be optimized during training to improve the response between different vertices in order to obtain more adaptive global information.

[0044] Please see Figure 5 In a specific embodiment of the present invention, the multi-scale information fusion unit includes a first branch and a second branch in parallel. The first branch is as follows: Figure 5 As shown on the left, it consists of a cascaded first convolutional layer, multiple parallel grouped convolutions, and a second convolutional layer; the second branch is as follows: Figure 5 As shown on the right, it consists of cascaded adaptive average pooling layers, a first convolutional layer, multiple parallel grouped convolutions, a second convolutional layer, and a first upsampling layer; finally, the output features of the first branch and the second branch are connected and then passed through the third convolutional layer and the second upsampling layer in sequence to obtain the output features of the multi-scale information fusion unit.

[0045] The first branch has a small receptive field and is responsible for targeting tiny objects. It first uses 1×1 convolutions to reduce the number of channels to 512, then aggregates several layers with different kernel sizes. The number of groups G makes the kernels have different connectivity. Each convolutional block is followed by a batch normalization layer and a ReLU activation layer.

[0046] The second branch has a similar structure to the first branch. In order to capture the features of large objects in the global perspective map, an adaptive average pooling operation is used at the top to reduce the spatial size of the feature map to 9×9, and bilinear interpolation is used at the bottom to upsample the feature map to the same resolution as the input.

[0047] After the two branches, the features output by the two branches are concatenated, followed by a standard convolutional layer of size 3×3. Finally, the feature maps are upsampled to the size of the original image. This process can improve the robustness of the model to scale changes.

[0048] In a specific embodiment of the present invention, the multiple parallel grouped convolutions include four grouped convolutions: the first grouped convolution has a kernel size of 9×9 and a group size of 16; the second grouped convolution has a kernel size of 7×7 and a group size of 8; the third grouped convolution has a kernel size of 5×5 and a group size of 4; and the fourth grouped convolution has a kernel size of 3×3 and a group size of 1.

[0049] In a specific embodiment of the present invention, a density map is generated by decoding through a deep convolutional structure. The process is as follows: In order to increase the resolution of the feature map, an upsampling operation is adopted; by using layer-by-layer upsampling and convolution, the image size is maintained and the number of channels is reduced, which not only preserves high-level semantic features, but also gradually restores spatial resolution. This multi-layer convolutional decoder not only has a powerful nonlinear mapping capability, but also can perform information fusion at multiple scales, thereby improving the accuracy and robustness of the decoding process, especially in the accurate identification and counting of target objects in complex environments.

[0050] In a specific embodiment of the present invention, in order to generate a more accurate density map and improve the accuracy of invert point prediction, the loss function during the training of the invert point detection network is as follows: K Overall =K Bayesian +λK Count Among them, K Overall K is the total loss function. Bayesian For Bayesian uncertainty loss, K Count Let λ be the counting error loss, and λ be a hyperparameter used to balance the weights of the two losses. A reliable supervision method improved by utilizing Bayesian uncertainty loss and counting error loss is used to learn the density probability, and then the expected count at each annotation is calculated, which can alleviate the non-uniformity of the density distribution to some extent.

[0051] To verify the effectiveness of the method in this invention, the inflection detection network in this invention was compared with other mainstream networks. The mean error (MAE) and root mean square error (RMSE) were used to evaluate the performance of the proposed method during the comparison. The comparison results are shown in the table below.

[0052] Table 1. Comparison of the performance of the network of this invention with various other networks on the same dataset.

[0053]

[0054] As can be seen from the table above, the inflection point detection network proposed in this invention outperforms mainstream networks in both MAE and RMSE, demonstrating excellent prediction accuracy. It is suitable for tasks on palm vein image datasets and shows absolute advantage among multiple traditional and emerging networks.

[0055] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A palm vein fold point detection method of adaptive noise suppression cooperative coupling feature enhancement, characterized by, The method comprises the following steps: obtaining a palm vein image to be processed; inputting the palm vein image into a trained fold point detection network to obtain the number and position information of palm vein fold points, the fold point detection network comprising a cascade of: a shallow feature extraction module for preprocessing the palm vein image; an adaptive noise suppression module for suppressing background noise to improve the signal-to-noise ratio of the image; a coupled feature enhancement module for enhancing the position information between each channel to obtain multi-scale information; and a decoder for processing the enhanced features to obtain the number and position information of the palm vein fold points. The adaptive noise suppression module processes the shallow features output by the shallow feature extraction module in the following steps: The coupled feature enhancement unit comprises a learnable graph neural unit and a multi-scale information fusion unit, the learnable graph neural unit uses a graph structure to enhance the spatial dependency between each channel, and the multi-scale information fusion unit uses a double-oriented pyramid operation to capture multi-scale information. dividing the shallow feature X into G sub-features X i , i ∈ {1, 2, …, G}; The sub-feature X is processed by using a GAP algorithm to obtain a first feature X i 1i ;​ using a constant ε on the first feature X 1i normalizing to obtain a second feature X 2i ; using learnable parameters γ and β on the second feature X 2i transformed to obtain a third feature X 3i ; Utilizing the sub-feature X i and the Sigmoid function to process the third feature X 3i , a fourth feature X 4i is obtained. stitching the fourth feature X 4i , obtaining an output feature Y of the adaptive noise suppression unit; The expression of the learnable graph neural unit is as follows: Z=Y·ReLU(Conv1d(A·GAP(Y))); A=A0×A1+A2; A0=Softmax(Conv1d(GAP(Y))); wherein Z is the output feature of the learnable graph neural unit, Y is the output feature of the adaptive noise suppression unit, GAP is a GAP algorithm, Conv1d is a one-dimensional convolution layer, ReLU is an activation function, Softmax is a normalization, A0, A1 and A2 are learnable adjacency matrices; The multi-scale information fusion unit comprises a first branch and a second branch in parallel, the first branch is composed of a cascade of a first convolution layer, a plurality of parallel grouped convolutions, and a second convolution layer, the second branch is composed of a cascade of an adaptive average pooling layer, a first convolution layer, a plurality of parallel grouped convolutions, a second convolution layer, and a first upsampling layer, and the output features of the first branch and the second branch are connected and then sequentially passed through a third convolution layer and a second upsampling layer to obtain the output feature of the multi-scale information fusion unit. The relationship between the features is expressed by the following formula:

2. The palm vein fold point detection method of claim 1, wherein, The plurality of parallel grouped convolutions comprises four grouped convolutions: X 1i =X i • GAP(X i ); ； X 3i =γX 2i +β; X 4i =X i • Sigmoid(X 3i ); Y = X 41 + X 42 +... + X 4G . 3.The palm vein fold point detection method of adaptive noise suppression and cooperative coupling feature enhancement according to claim 1, wherein, The first grouped convolution has a kernel size of 9x9 and a group number of 16; The second grouped convolution has a kernel size of 7x7 and a group number of 8; The third grouped convolution has a kernel size of 5x5 and a group number of 4; The fourth grouped convolution has a kernel size of 3x3 and a group number of 1. The loss function of the fold point detection network during training is as follows:

4. The method of claim 1, wherein the adaptive noise suppression cooperatively coupled feature enhancement of palm vein fold point detection is characterized by, ​ K Overall =K Bayesian +λK Count ; where K Overall is the total loss function, K Bayesian is the Bayesian uncertainty loss, K Count is the count error loss, and λ is a hyperparameter used to balance the weight of the two losses.

Images