Image recognition method based on bionic vision and fluctuation enhancement

Abstract

The invention belongs to the technical field of image recognition, and particularly relates to an image recognition method based on bionic vision and fluctuation enhancement, and the method comprises the steps: building a bionic vision network model based on a multi-scale Gabor filter bank and a similarity measurement-based adaptive enhancement mechanism; establishing a pre-training module based on self-supervised learning, and pre-training the bionic vision network model; establishing a visual dynamic attention module, adding the visual dynamic attention module into the bionic visual network model, performing fluctuation enhancement and attention weighting on output features of the bionic visual network model, and fusing feature representations after fluctuation enhancement and attention weighting; combining supervision comparison loss, balance logarithm loss and focus loss to form a comprehensive loss function, and training the bionic vision network model; and performing image recognition by using the trained bionic visual network model. According to the method, the accuracy and robustness of image recognition and the adaptability to complex scenes are remarkably improved.

Description

Technical Field

[0001] The invention belongs to the technical field of image recognition, and in particular relates to an image recognition method based on bionic vision and wave enhancement. Background Art

[0002] In the current field of image recognition technology, deep learning models, especially convolutional neural networks (CNNs), have become mainstream and are widely used in many fields such as computer vision, medical imaging, and autonomous driving. These models effectively extract image features through deep structures and achieve high recognition accuracy on simple images. However, the recognition accuracy of existing algorithms is often limited when processing images with complex textures and details, especially under conditions of lighting changes, noise interference, and complex backgrounds, where their performance decreases significantly. In addition, deep learning models usually require a lot of computing resources, which limits their application in resource-constrained environments.

[0003] The main shortcomings of existing technologies include: insufficient image feature extraction, insufficient ability to capture high-frequency information and fine structures; insufficient sensitivity of the algorithm to image edges and local features, resulting in reduced recognition accuracy in complex environments; high computational costs of deep network structures limit their application in resource-constrained environments; insufficient robustness, and susceptibility to external interference, leading to recognition errors. These problems show that although existing technologies have made significant progress in the field of image recognition, further innovation and improvement are still needed to improve recognition accuracy and robustness. Summary of the invention

[0004] In view of the above deficiencies in the prior art, the purpose of the present invention is to provide an image recognition method based on bionic vision and wave enhancement, which significantly improves the accuracy, robustness and adaptability of image recognition to complex scenes.

[0005] To achieve the above objectives, the present invention provides an image recognition method based on bionic vision and wave enhancement, comprising the following steps: S1. Establish a bionic visual network model based on multi-scale Gabor filter banks and an adaptive enhancement mechanism based on similarity measurement; S2. Establish a pre-training module based on self-supervised learning to pre-train the bionic visual network model; S3. Establish a visual dynamic attention module and add it to the bionic visual network model, perform fluctuation enhancement and attention weighting on the output features of the bionic visual network model, and fuse the feature representations after fluctuation enhancement and attention weighting, wherein the visual dynamic attention module includes a fluctuation enhancement visual module, a dynamic attention mechanism and an attention feature fusion module; S4. Combine the supervised contrast loss, balanced logarithmic loss and focal loss to form a comprehensive loss function, and train the bionic visual network model based on the comprehensive loss function; S5. Use the trained bionic visual network model to perform image recognition.

[0006] As a preferred solution of the present invention, in S1, the multi-scale Gabor filter bank is expressed as: ; In the formula, , is the scale parameter of the filter in the x and y directions; is the direction of the filter; is the center frequency of the filter; the multi-scale Gabor filter bank is adjusted by , , and , generates filters of different scales and directions to capture multiple features in the image; x, y represent the two-dimensional coordinates of the input; j is a complex unit; For the input image , a multi-scale Gabor filter bank is used for feature extraction to obtain a set of multi-scale and multi-directional feature maps , expressed as: ; In the formula, Represents the Gabor filter function, which is a filter for scale m and direction n.

[0007] As a preferred solution of the present invention, in S1, the adaptive enhancement mechanism based on similarity measurement adopts the structural similarity index SSIM, which is expressed as: ; Where p and q represent two image blocks of comparative similarity; , are the grayscale means of image blocks p and q respectively; is the covariance of image blocks p and q, representing the joint distribution of grayscale; C1 and C2 are constants used to stabilize the calculation; Adaptive enhancement based on similarity measurement is achieved through the following steps: for , first in The local area is calculated with The structural similarity index SSIM is expressed as: ; In the formula, is the weight of the local area, reflecting and Similarities between corresponding regions; After obtaining the local weight of each feature map in the multi-scale and multi-directional feature map, the feature maps of different scales and directions are weighted fused to generate the adaptively enhanced feature representation. : .

[0008] As a preferred solution of the present invention, in S2, establishing a pre-training module based on self-supervised learning includes three steps, namely, defining an enhancement strategy, encoding and projection, and calculating loss and back propagation, specifically: S11. Define enhancement strategies, including spatial transformation strategies and temporal transformation strategies. The spatial transformation strategies include random cropping, resizing, flipping and rotation. The temporal transformation strategies simulate the degradation effects of images over time, including color distortion, Gaussian blur and random noise. For the unlabeled image dataset obtained, perform the spatial transformation strategy and the temporal transformation strategy on each image in the dataset to obtain enhanced images. and ; S12, encoding and projection, using the bionic visual network model as the encoder, and Perform feature extraction, and the obtained features are expressed as and , and use a multi-layer perceptron MLP for feature transformation and projection: ; ; In the formula, represents the MLP transformation, , Represents the transformed features; Through the projection head The transformed features are further processed to obtain the projected features and : ; ; S13. Calculate loss and back propagation, using contrastive learning loss function L c Calculate the loss and optimize the parameters of the bionic visual network model through back propagation, L c It is expressed as: ; Where Nd is the number of sample pairs used to calculate the loss, k = 1, 2, ..., Nd; , represents the contrast loss, which is defined as: ; ; In the formula, Indicates the similarity between features; is the temperature parameter; A represents the set of all negative samples.

[0009] As a preferred embodiment of the present invention, in S3, the fluctuation enhancement visual module is: Extract preliminary features from the input image through the bionic visual network model: ; In the formula, represents the bionic visual network model; for , the wave theory is applied to enhance the features of the image. Specifically, the two-dimensional Helmholtz equation is used to simulate the high-frequency features in the image, so that the edge and texture information of the image are enhanced, which is expressed as: ; In the formula, The characteristic map after fluctuation enhancement is shown; is the Laplace operator; t is the wave number; After fluctuation enhancement, the peripheral vision submodule The peripheral vision submodule dynamically adjusts the surrounding area to extract local information related to the global features. The query position, that is, the central pixel of the local area, is selected as the visual center. The peripheral local space of the central area is planned into four sub-areas: center, quasi-center, middle periphery, and far periphery. The query in the central area interacts with the keys of the four sub-areas, and the interaction intensity gradually decreases according to the change of distance. In the interaction process, the fluctuation enhancement distance is used. Matching interaction strength, fluctuation enhancement distance The wave equation is used to simulate the mutual influence between features, which is expressed as: ; Where N is the number of feature points, e=1, 2, ..., N; is the feature q e The wave solution of Enhanced distance based on fluctuations The feature output representation of the attention mechanism for: ; In the formula, and is a learnable weight matrix; is the activation function; is a nonlinear activation function; represents a linear activation operation.

[0010] As a preferred embodiment of the present invention, in S3, the dynamic attention mechanism is: for , through linear transformation Mapping to query vector Q, key vector K and value vector V: ; ; ; Where W Q , W K and W V is a learnable weight matrix; calculate the similarity between Q and K, and normalize it through the softmax function to get the attention weight; ; in is the attention weight between positions i and j; is the square root of the dimension of the key vector; Q i is the position i and is the query vector; K j is the key vector at position j; T denotes transpose; According to the calculated , perform weighted summation on V to generate enhanced feature representation: ; In the formula, O i is the output feature representation after enhancement at position i; V j is the value vector at position j; Summarize the enhanced output feature representations of all positions i to obtain the attention-weighted feature representation O, which is O={O1, O2, …, O i}.

[0011] As a preferred solution of the present invention, in S3, the attention feature fusion module is: The output of the fluctuation enhanced vision module The output feature representation enhanced by the dynamic attention mechanism is fused to further optimize the feature representation. for: ; Where α and β are learnable weight parameters used to balance the contribution of the fluctuation enhancement distance and the attention mechanism output in the final fusion feature; O is the attention-weighted feature representation calculated by the dynamic attention mechanism.

[0012] As a preferred embodiment of the present invention, in S4, the process of establishing the comprehensive loss function is: Supervised Contrastive Loss It is expressed as: ; Where Nd is the number of sample pairs used to calculate the loss, a=1, 2, ..., Nd; u a 、u b 、u c They represent the features extracted by the bionic visual network model for the ath, bth, and cth samples respectively; Represents the positive sample set to which the a-th sample belongs; Represents the number of a-th positive samples in the data set; is the temperature parameter; T represents transposition; Balanced Log Loss It is expressed as: ; In the formula, y a is the true label of the ath sample, which is 0 or 1; p a is the probability that the model predicts that the ath sample belongs to category 1; w a is the sample weight of the ath sample; Focal Loss It is expressed as: ; In the formula, represents the difficulty weight; is the focus parameter; Combine balanced logarithmic loss, focal loss and supervised contrast loss to form a comprehensive loss function , expressed as: ; In the formula, Indicates u a with u b The similarity measure between .

[0013] As a preferred embodiment of the present invention, in S5, the specific process of image recognition is as follows: the input image After being processed by a multi-scale Gabor filter bank and an adaptive enhancement mechanism, the adaptively enhanced feature representation is obtained. , and then through the fluctuation enhanced visual module, the feature output representation is obtained At the same time, through the dynamic attention mechanism The attention-weighted feature representation O is calculated, and finally the fused feature representation is obtained through the attention feature fusion module. , which is the final optimized feature representation, completing the recognition of the image.

[0014] The model and algorithm involved in the present invention can be executed by an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the software to implement the above-mentioned model and algorithm calculation.

[0015] The beneficial effects of the present invention are: The present invention has a number of significant advantages through an image recognition algorithm that integrates bionic visual processing and wave theory. First, by simulating the primary visual processing of the human retina, the Gabor filter group is used to perform multi-scale and multi-directional filtering operations to effectively extract the edge and texture information of the image; this bionic visual preprocessing network is combined with self-supervised learning for pre-training, which enhances the model's generalization ability for unlabeled data and improves its sensitivity to details and texture information in complex image scenes. Secondly, the physical wave theory introduced enhances image features through the two-dimensional Helmholtz equation, captures more complex edge and texture features, and significantly improves the accuracy of image recognition.

[0016] The present invention uses a dynamic visual attention module, combined with fluctuation enhancement and attention mechanism, to adaptively adjust the attention weight, so that the model can focus on the key areas in the image more accurately. This mechanism not only improves the effect of feature extraction, but also enhances the adaptability and robustness of the model to complex image scenes.

[0017] The present invention further optimizes the training process of the model by combining balanced logarithmic loss, focal loss and supervised contrast loss when designing the comprehensive loss function, especially when dealing with unbalanced data sets and difficult-to-classify samples, thereby improving the classification accuracy and robustness of the model. It has broad application potential in high-demand application scenarios such as autonomous driving, medical image analysis, and remote sensing image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is the overall process principle diagram of the present invention; Figure 2 It is a flow chart of the image recognition process of the present invention. DETAILED DESCRIPTION

[0019] The embodiments of the present invention are further described below in conjunction with the accompanying drawings: like Figure 1As shown, an image recognition method based on bionic vision and wave enhancement includes the following steps: S1. Establish a bionic visual network model based on multi-scale Gabor filter banks and an adaptive enhancement mechanism based on similarity measurement; S2. Establish a pre-training module based on self-supervised learning to pre-train the bionic visual network model; S3. Establish a visual dynamic attention module and add it to the bionic visual network model, perform fluctuation enhancement and attention weighting on the output features of the bionic visual network model, and fuse the feature representations after fluctuation enhancement and attention weighting, wherein the visual dynamic attention module includes a fluctuation enhancement visual module, a dynamic attention mechanism and an attention feature fusion module; S4. Combine the supervised contrast loss, balanced logarithmic loss and focal loss to form a comprehensive loss function, and train the bionic visual network model based on the comprehensive loss function; S5. Use the trained bionic visual network model to perform image recognition.

[0020] In S1, the multi-scale Gabor filter bank is expressed as: ; In the formula, , is the scale parameter of the filter in the x and y directions; is the direction of the filter; is the center frequency of the filter; the multi-scale Gabor filter bank is adjusted by , , and , generates filters of different scales and directions to capture multiple features in the image; x, y represent the two-dimensional coordinates of the input; j is a complex unit; For the input image , a multi-scale Gabor filter bank is used for feature extraction to obtain a set of multi-scale and multi-directional feature maps , expressed as: ; In the formula, Represents the Gabor filter function, which is a filter for scale m and direction n.

[0021] Since the dynamic response characteristics of the Gabor function are very consistent with the physiological characteristics of the human visual system, it has a certain correlation. At the same time, it also has strong reliability and stability to environmental changes and expression changes, and has excellent performance in spatial locality and selectivity in different directions. Gabor filters are widely used to simulate the primary visual cortex function in the human visual system. Therefore, a multi-scale and multi-directional Gabor filter group is designed to extract different features in the image.

[0022] In order to further improve the recognition ability of features, a frequency domain separation mechanism can be used, especially to separate high-frequency information and low-frequency information. The frequency domain information of the image is extracted through non-sampled shearlet transform (NSCT): ; In the formula, and Represent the low-frequency and high-frequency components of image I respectively.

[0023] In S1, the adaptive enhancement mechanism based on similarity measurement adopts the structural similarity index SSIM, which is expressed as: ; Where p and q represent two image blocks of comparative similarity; , are the grayscale means of image blocks p and q respectively; is the covariance of image blocks p and q, representing the joint distribution of grayscale; C1 and C2 are constants used to stabilize the calculation; Adaptive enhancement based on similarity measurement is achieved through the following steps: for , first in The local area is calculated with The structural similarity index SSIM is expressed as: ; In the formula, is the weight of the local area, reflecting and The similarity between corresponding regions; when it approaches 1, it indicates that the structure of the extracted feature is very similar to that of the original image, so this feature should be given a higher weight in the fusion; After obtaining the local weight of each feature map in the multi-scale and multi-directional feature map, the feature maps of different scales and directions are weighted fused to generate the adaptively enhanced feature representation. : .

[0024] In the feature fusion process of the bionic visual network, in order to more effectively focus on the key areas in the image, an adaptive enhancement mechanism based on similarity measurement is introduced. This mechanism dynamically adjusts the weight of feature fusion to enhance the adaptability and robustness of the model when processing complex image scenes. It not only considers the information of various scales and directions, but also dynamically adjusts the weights of different areas, so that the features of the key areas occupy a larger proportion in the fusion results.

[0025] The Structural Similarity Index (SSIM) is a measure of the perceived similarity between two images. SSIM not only considers the difference between pixels, but also considers higher-level image information such as brightness, contrast, and structure. The value calculated by SSIM is between [0, 1], and the larger the value, the higher the similarity between the two image blocks.

[0026] In S2, establishing a pre-training module based on self-supervised learning includes three steps, namely, defining the enhancement strategy, encoding and projection, and calculating the loss and back propagation, specifically: S11. Define enhancement strategies, including spatial transformation strategies and temporal transformation strategies. The spatial transformation strategies include random cropping, resizing, flipping and rotation. The temporal transformation strategies simulate the degradation effects of images over time, including color distortion, Gaussian blur and random noise. For the unlabeled image dataset obtained, perform the spatial transformation strategy and the temporal transformation strategy on each image in the dataset to obtain enhanced images. and ; S12, encoding and projection, using the bionic visual network model as the encoder, and Perform feature extraction, and the obtained features are expressed as and , and use a multi-layer perceptron MLP for feature transformation and projection: ; ; In the formula, represents the MLP transformation, , Represents the transformed features; Through the projection head The transformed features are further processed to obtain the projected features and : ; ; S13. Calculate loss and back propagation, using contrastive learning loss function L cCalculate the loss and optimize the parameters of the bionic visual network model through back propagation, L c It is expressed as: ; Where Nd is the number of sample pairs used to calculate the loss, k = 1, 2, ..., Nd; , represents the contrast loss, which is defined as: ; ; In the formula, Indicates the similarity between features; is the temperature parameter; A represents the set of all negative samples.

[0027] Contrastive Learning is a self-supervised learning method. The main feature of self-supervised learning is that it does not require manually annotated data labels, but instead generates labels from the data itself by designing pretext tasks for training. Contrastive learning generates positive and negative sample pairs, aiming to make the feature representations of positive sample pairs more similar and the feature representations of negative sample pairs more dissimilar, thereby learning useful feature representations. Using the contrastive learning loss function L c The loss is computed to force the feature representations of two augmented versions of the same image to be close, and the feature representations of different images to be far apart.

[0028] In S3, the fluctuation enhanced vision module is: Extract preliminary features from the input image through the bionic visual network model: ; In the formula, represents the bionic visual network model; for , the wave theory is applied to enhance the features of the image. Specifically, the two-dimensional Helmholtz equation is used to simulate the high-frequency features in the image, so that the edge and texture information of the image are enhanced, which is expressed as: ; In the formula, The characteristic map after fluctuation enhancement is shown; is the Laplace operator; t is the wave number; After fluctuation enhancement, the peripheral vision submodule The peripheral vision submodule dynamically adjusts the surrounding area to extract local information related to the global features. The query position, that is, the central pixel of the local area, is selected as the visual center. The peripheral local space of the central area is planned into four sub-areas: center, quasi-center, middle periphery, and far periphery. The query in the central area interacts with the keys of the four sub-areas, and the interaction intensity gradually decreases according to the change of distance. In the interaction process, the fluctuation enhancement distance is used. Matching interaction strength (can better reflect the enhancement effect based on wave theory), wave enhancement distance The wave equation is used to simulate the mutual influence between features, which is expressed as: ; Where N is the number of feature points, e=1, 2, ..., N; is the feature q e The wave solution of Able to capture more complex feature changes in images; Enhanced distance based on fluctuations The feature output representation of the attention mechanism for: ; In the formula, and is a learnable weight matrix; is the activation function; is a nonlinear activation function; represents a linear activation operation.

[0029] Among them, the peripheral vision module can be constructed and implemented based on the context perception model, regional attention mechanism and edge detection algorithm in the known technology. The fluctuation enhanced distance can more accurately measure the relationship between the query vector and the key vector by considering the characteristics of fluctuation propagation, so as to better capture the key areas in the image in the attention mechanism. The fluctuation enhanced distance replaces the traditional Euclidean distance, and the attention mechanism is more adaptable to the changes of complex image features in combination with the fluctuation theory, thereby improving the performance of the model in image recognition tasks.

[0030] In S3, the dynamic attention mechanism is: for , through linear transformation Mapping to query vector Q, key vector K and value vector V: ; ; ; Where W Q , WK and W V is a learnable weight matrix; calculate the similarity between Q and K, and normalize it through the softmax function to get the attention weight; ; in is the attention weight between positions i and j; is the square root of the dimension of the key vector; Q i is the position i and is the query vector; K j is the key vector at position j; T denotes transpose; According to the calculated , perform weighted summation on V to generate enhanced feature representation: ; In the formula, O i is the output feature representation after enhancement at position i; V j is the value vector at position j; Summarize the enhanced output feature representations of all positions i to obtain the attention-weighted feature representation O, which is O={O1, O2, …, O i}.

[0031] The attention mechanism is a technology that enables neural networks to focus on the most relevant parts of data when processing data. It was first proposed and applied in the field of natural language processing (NLP), and later gradually expanded to other fields such as image processing. Attention mechanisms can be divided into many types, such as self-attention, multi-head attention, etc. The dynamic attention module is an adaptive attention mechanism, which mainly consists of attention calculation, relevance weight, feature aggregation, and adaptive adjustment mechanism. It can dynamically adjust the attention weight according to the changes in input features, enhance the feature extraction ability of the neural network, and enable it to adaptively adjust the attention weights between different pixels in the image, so as to more effectively capture global information. It is suitable for processing complex image classification tasks, especially on unbalanced data sets.

[0032] In S3, the attention feature fusion module is: The output of the fluctuation enhanced vision module The output feature representation enhanced by the dynamic attention mechanism is fused to further optimize the feature representation. for: ; Where α and β are learnable weight parameters used to balance the contribution of the fluctuation enhancement distance and the attention mechanism output in the final fusion feature; O is the attention-weighted feature representation calculated by the dynamic attention mechanism.

[0033] In S4, the process of establishing the comprehensive loss function is: Supervised contrastive learning increases the number and labels of positive samples on the basis of self-supervised learning, and establishes a connection between contrastive loss and the true distribution of data, which can help the neural network extract more reasonable features and improve the relevance of contrastive learning and downstream tasks. It is expressed as: ; Where Nd is the number of sample pairs used to calculate the loss, a=1, 2, ..., Nd; u a 、u b 、u c They represent the features extracted by the bionic visual network model for the ath, bth, and cth samples respectively; Represents the positive sample set to which the a-th sample belongs; Represents the number of a-th positive samples in the data set; is the temperature parameter; T represents transposition; Balanced log loss is an improved cross entropy loss used to deal with the problem of imbalanced categories. It introduces category weights to ensure that samples of each category receive appropriate attention in loss calculation, which is suitable for classification tasks with an imbalanced number of samples. Balanced Log Loss It is expressed as: ; In the formula, y a is the true label of the ath sample, which is 0 or 1; p a is the probability that the model predicts that the ath sample belongs to category 1; w a is the sample weight of the ath sample; by assigning higher weights to minority class samples, the model will pay more attention to these samples during training, thereby effectively dealing with the problem of class imbalance, and allowing weights to be manually set according to the characteristics of the dataset and specific problems, making the model more adaptable to different degrees of class imbalance; Focal Loss It is expressed as: ; In the formula, Represents the difficulty weight, which can increase the weight of difficult-to-classify samples; is the focusing parameter. When , the focal loss degenerates into the ordinary cross entropy loss; Combine balanced logarithmic loss, focal loss and supervised contrast loss to form a comprehensive loss function , expressed as: ; In the formula, Indicates u a with u b The similarity measure between them. By introducing category weights, the contribution of minority class samples to the loss is increased, thereby guiding the model to better learn the features of these categories. At the same time, combined with focal loss, the weight of difficult-to-classify samples is increased, so that the model pays more attention to difficult-to-classify samples during training, improves classification accuracy, and enhances the feature representation effect of the model, making the model more accurate and robust when processing image classification tasks.

[0034] like Figure 2 As shown in S5, the specific process of image recognition is as follows: the input image After being processed by a multi-scale Gabor filter bank and an adaptive enhancement mechanism, the adaptively enhanced feature representation is obtained. , and then through the fluctuation enhanced visual module, the feature output representation is obtained At the same time, through the dynamic attention mechanism The attention-weighted feature representation O is calculated, and finally the fused feature representation is obtained through the attention feature fusion module. , which is the final optimized feature representation, completing the recognition of the image.

Claims

1. An image recognition method based on bionic vision and wave enhancement, Features The following steps are involved: S1. Establish a bionic visual network model based on multi-scale Gabor filter banks and an adaptive enhancement mechanism based on similarity measurement; S2. Establish a pre-training module based on self-supervised learning to pre-train the bionic visual network model; S3. Establish a visual dynamic attention module and add it to the bionic visual network model, perform fluctuation enhancement and attention weighting on the output features of the bionic visual network model, and fuse the feature representations after fluctuation enhancement and attention weighting, wherein the visual dynamic attention module includes a fluctuation enhancement visual module, a dynamic attention mechanism and an attention feature fusion module; S4. Combine the supervised contrast loss, balanced logarithmic loss and focal loss to form a comprehensive loss function, and train the bionic visual network model based on the comprehensive loss function; S5. Use the trained bionic visual network model to perform image recognition.

2. According to claim 1, an image recognition method based on bionic vision and wave enhancement, Features: In S1, the multi-scale Gabor filter bank is expressed as: ； In the formula, , is the scale parameter of the filter in the x and y directions; is the direction of the filter; is the center frequency of the filter; the multi-scale Gabor filter bank is adjusted by , , and , generate filters of different scales and orientations to capture multiple features in the image; x, y represent the two-dimensional coordinates of the input; j is the complex unit; For the input image , a multi-scale Gabor filter bank is used for feature extraction to obtain a set of multi-scale and multi-directional feature maps , expressed as: ； In the formula, Represents the Gabor filter function, which is a filter for scale m and direction n.

3. According to claim 2, an image recognition method based on bionic vision and wave enhancement, Features: In S1, the adaptive enhancement mechanism based on similarity measurement adopts the structural similarity index SSIM, which is expressed as: ； Where p and q represent two image blocks of comparative similarity; , are the grayscale means of image blocks p and q respectively; is the covariance of image blocks p and q, indicating the joint distribution of grayscale; C 1 and C 2 is a constant used to stabilize the calculation; Adaptive enhancement based on similarity measurement is achieved through the following steps: for , first in The local area is calculated with The structural similarity index SSIM is expressed as: ； In the formula, is the weight of the local area, reflecting and Similarities between corresponding regions; After obtaining the local weight of each feature map in the multi-scale and multi-directional feature map, the feature maps of different scales and directions are weighted fused to generate the adaptively enhanced feature representation. : 。 4. According to claim 3, an image recognition method based on bionic vision and wave enhancement, Features: In S2, establishing a pre-training module based on self-supervised learning includes three steps, namely, defining an enhancement strategy, encoding and projection, and calculating loss and back propagation, specifically: S11. Define enhancement strategies, including spatial transformation strategies and temporal transformation strategies. The spatial transformation strategies include random cropping, resizing, flipping and rotation. The temporal transformation strategies simulate the degradation effects of images over time, including color distortion, Gaussian blur and random noise. For the unlabeled image dataset obtained, perform the spatial transformation strategy and the temporal transformation strategy on each image in the dataset to obtain enhanced images. and ; S12, encoding and projection, using the bionic visual network model as the encoder, and Perform feature extraction, and the obtained features are expressed as and , and use a multi-layer perceptron MLP for feature transformation and projection: ； ； In the formula, represents the MLP transformation, , Represents the transformed features; Through the projection head The transformed features are further processed to obtain the projected features and : ； ； S13. Calculate loss and back propagation, using contrastive learning loss function L c Calculate the loss and optimize the parameters of the bionic visual network model through back propagation, L c It is expressed as: ； Where Nd is the number of sample pairs used to calculate the loss, k = 1, 2, ..., Nd; , represents the contrast loss, which is defined as: ； ； In the formula, Indicates the similarity between features; is the temperature parameter; A represents the set of all negative samples.

5. According to claim 4, an image recognition method based on bionic vision and wave enhancement, Features: In the S3, the fluctuation enhanced visual module is: Extract preliminary features from the input image through the bionic visual network model: ； In the formula, represents the bionic visual network model; for , the wave theory is applied to enhance the features of the image. Specifically, the two-dimensional Helmholtz equation is used to simulate the high-frequency features in the image, so that the edge and texture information of the image are enhanced, which is expressed as: ； In the formula, The characteristic map after fluctuation enhancement is shown; is the Laplace operator; t is the wave number; After fluctuation enhancement, the peripheral vision submodule The peripheral vision submodule dynamically adjusts the surrounding area to extract local information related to the global features. The query position, that is, the central pixel of the local area, is selected as the visual center. The peripheral local space of the central area is planned into four sub-areas: center, quasi-center, middle periphery, and far periphery. The query in the central area interacts with the keys of the four sub-areas, and the interaction intensity gradually decreases according to the change of distance. In the interaction process, the fluctuation enhancement distance is used. Matching interaction strength, fluctuation enhancement distance The wave equation is used to simulate the mutual influence between features, which is expressed as: ； Where N is the number of feature points, e=1, 2, ..., N; is the feature q e The wave solution of Enhanced distance based on fluctuations The feature output representation of the attention mechanism for: ； In the formula, and is a learnable weight matrix; is the activation function; is a nonlinear activation function; represents a linear activation operation.

6. The image recognition method based on bionic vision and wave enhancement according to claim 5, Features: In S3, the dynamic attention mechanism is: for , through linear transformation Mapping to query vector Q, key vector K and value vector V: ； ； ； Where W Q , W K and W V is a learnable weight matrix; calculate the similarity between Q and K, and normalize it through the softmax function to get the attention weight; ； in is the attention weight between positions i and j; is the square root of the dimension of the key vector; Q i is the position i and is the query vector; K j is the key vector at position j; T denotes transpose; According to the calculated , perform weighted summation on V to generate enhanced feature representation: ； In the formula, O i It is the output feature representation after enhancement at position i; V j is the value vector at position j; Summarize the enhanced output feature representations of all positions i to obtain the attention-weighted feature representation O, which is O={O 1 , O 2 , ..., O i }.

7. The image recognition method based on bionic vision and wave enhancement according to claim 6, Features: In the above S3, the attention feature fusion module is: The output of the fluctuation enhanced vision module The output feature representation enhanced by the dynamic attention mechanism is fused to further optimize the feature representation. for: ； Where α and β are learnable weight parameters used to balance the contribution of the fluctuation enhancement distance and the attention mechanism output in the final fusion feature; O is the attention-weighted feature representation calculated by the dynamic attention mechanism.

8. The image recognition method based on bionic vision and wave enhancement according to claim 1, Features: In the above S4, the process of establishing the comprehensive loss function is: Supervised Contrastive Loss It is expressed as: ； Where Nd is the number of sample pairs used to calculate the loss, a=1, 2, ..., Nd; u a 、u b 、u c They represent the features extracted by the bionic visual network model for the ath, bth, and cth samples respectively; Represents the positive sample set to which the a-th sample belongs; Represents the number of a-th positive samples in the data set; is the temperature parameter; T represents transposition; Balanced Log Loss It is expressed as: ； In the formula, y a is the true label of the ath sample, which is 0 or 1; p a is the probability that the model predicts that the ath sample belongs to category 1; w a is the sample weight of the a-th sample; Focal Loss It is expressed as: ； In the formula, represents the difficulty weight; is the focus parameter; Combine balanced logarithmic loss, focal loss and supervised contrast loss to form a comprehensive loss function , expressed as: ； In the formula, Indicates u a with u b The similarity measure between .

9. The image recognition method based on bionic vision and wave enhancement according to claim 7, Features: In S5, the specific process of image recognition is as follows: the input image After being processed by a multi-scale Gabor filter bank and an adaptive enhancement mechanism, the adaptively enhanced feature representation is obtained. , and then through the fluctuation enhanced visual module, the feature output representation is obtained At the same time, through the dynamic attention mechanism The attention-weighted feature representation O is calculated, and finally the fused feature representation is obtained through the attention feature fusion module. , which is the final optimized feature representation, completing the recognition of the image.

Images