Enhancement processing method and system for zebrafish behavior image

Abstract

This invention relates to the field of image enhancement processing technology, specifically including a method and system for enhancing zebrafish behavior images. The method includes: inputting a set of zebrafish images, training a dual-branch enhancement adaptation model, analyzing the first and second types of zebrafish image sets, and obtaining a set of zebrafish behavior-enhanced images. It solves the technical problems that uniform enhancement parameters cannot match the differences in illumination and scattering between shallow and deep water, that the enhancement process easily leads to feature distortion or loss of details, that it is difficult to comprehensively capture multi-scale behavioral details, and that the behavioral feature recognition of the enhanced image is insufficient. It achieves differentiated enhancement of shallow-water and deep-water zebrafish images, adapts to the scene characteristics of different water layers, improves the accuracy of behavioral feature extraction in shallow and deep water scenes, ensures the integrity of behavioral features while improving image quality, enhances the ability to capture multi-scale behavioral details and the recognition of behavioral features, and ensures the accuracy of zebrafish behavior analysis.

Description

Technical Field

[0001] This invention relates to the field of image enhancement processing technology, specifically to a method and system for enhancing zebrafish behavioral images. Background Technology

[0002] Zebrafish often move in different water layers, and the lighting conditions and water scattering characteristics of different water layers are significantly different. This leads to problems such as uneven brightness, insufficient contrast, and blurred details in the collected zebrafish images, which seriously affects the accuracy of subsequent behavioral feature extraction and analysis.

[0003] Current methods for enhancing zebrafish behavior images often use uniform enhancement parameters, failing to fully consider the differences in shallow and deep water layers where zebrafish are active. They also ignore the differences in image features caused by illumination and scattering at different water layers, making it difficult to accurately capture the differences in behavioral features in shallow and deep water scenes. As a result, behavioral feature distortion or loss of details can easily occur after enhancement, which restricts the accuracy of zebrafish behavior analysis.

[0004] In summary, existing technologies suffer from several technical problems: uniform enhancement parameters cannot match the differences in lighting and scattering between shallow and deep water; the enhancement process is prone to feature distortion or loss of detail; it is difficult to fully capture multi-scale behavioral details; and the behavioral feature recognition of the enhanced image is insufficient. Summary of the Invention

[0005] This application provides a method and system for enhancing zebrafish behavior images, aiming to solve the technical problems in the prior art where uniform enhancement parameters cannot match the differences in illumination and scattering between shallow and deep water, the enhancement process is prone to feature distortion or loss of details, it is difficult to fully capture multi-scale behavioral details, and the behavioral feature recognition of the enhanced image is insufficient.

[0006] In view of the above problems, the technical solution to achieve the present application is as follows:

[0007] In a first aspect, this application provides a method for enhancing zebrafish behavior images. The method includes: inputting a set of zebrafish images; extracting water layer-related features from each zebrafish image in the set; dividing the set of zebrafish images into a first-class set of zebrafish images with shallow-water labels and a second-class set of zebrafish images with deep-water labels according to the water layer-related features; training a two-branch enhancement adaptation model based on a pre-constructed first-class representative behavior feature sample library and a second-class representative behavior feature sample library, wherein the two-branch enhancement adaptation model includes a shared feature extraction network, a two-branch feature multi-layer convolutional network, and a two-branch enhancement parameter adaptation network; and analyzing the first-class and second-class zebrafish image sets using the two-branch enhancement adaptation model, and processing the first-class and second-class zebrafish image sets according to the obtained first-class and second-class enhancement processing parameters to obtain a set of zebrafish behavior-enhanced images.

[0008] Preferably, water-layer related features are extracted from each zebrafish image. These water-layer related features include at least global average brightness features, histogram distribution features, background local binary variance, background texture variance features, and background edge distribution features. A training dataset is constructed, comprising a first-class zebrafish training image set with known shallow-water labels and a second-class zebrafish training image set with known deep-water labels, as well as corresponding first-group and second-group water-layer related feature training samples. A support vector machine (SVM) is used as a base classifier to train a classification model on the training dataset, resulting in an SVM model with a classification accuracy greater than a preset accuracy. The converged SVM model performs label classification of the zebrafish image set based on the input water-layer related features.

[0009] Preferably, a radial basis function kernel is introduced, which is obtained by configuring a 6-dimensional feature dimension, the feature importance weights corresponding to the 6-dimensional feature dimension, and the feature standard deviation; hyperparameters are initialized, including penalty parameters and kernel function parameters; the SMO optimization algorithm is used to train the model on the standardized training dataset to obtain Lagrange multipliers and biases, and the support vector machine trained based on the Lagrange multipliers and biases is used for model performance testing and optimization until a support vector machine model with a classification accuracy greater than the preset accuracy is obtained.

[0010] Preferably, the shared feature extraction network extracts a first set of feature vectors and a second set of feature vectors for the first and second types of zebrafish image sets. The first set of feature vectors is sent to the first branch of a dual-branch feature multi-layer convolutional network, and the second set of feature vectors is sent to the second branch of the dual-branch feature multi-layer convolutional network for multi-scale feature convolution, resulting in shallow water and deep water feature vector sets. The dual-branch enhancement parameter adaptation network optimizes the shallow water feature vector set in the enhancement parameter action space using a defined first feature similarity loss function, and optimizes the deep water feature vector set in the enhancement parameter action space using a defined second feature similarity loss function, resulting in first and second types of enhancement processing parameters. The enhancement parameter action space includes brightness parameters, contrast parameters, saturation parameters, sharpening intensity parameters, color balance parameters, and noise suppression intensity parameters.

[0011] Preferably, the shared feature extraction network is obtained through training by connecting three layers of neural networks. The three layers of neural networks include pixel feature extraction in a first interval, pixel feature extraction in a second interval, and pixel feature extraction in a third interval; the first interval is 52×52, the second interval is 26×26, and the third interval is 13×13.

[0012] Preferably, the first branch feature multilayer convolutional network is constructed from K convolutional layers consisting of K first convolutional kernel clustering groups, and the second branch feature multilayer convolutional network is constructed from K convolutional layers consisting of K second convolutional kernel clustering groups; wherein, the K first convolutional kernel clustering groups are convolutional kernels obtained by selecting the first feature pixel scale set belonging to the first class representative behavior feature sample library and performing K-value clustering on the first feature pixel scale set; the K second convolutional kernel clustering groups are convolutional kernels obtained by selecting the second feature pixel scale set belonging to the second class representative behavior feature sample library and performing K-value clustering on the first feature pixel scale set.

[0013] Preferably, an initial enhancement parameter strategy set is obtained in the enhancement parameter action space; the shallow water behavior feature vector set is enhanced according to the initial enhancement parameter strategy set, the enhanced shallow water behavior feature vector set is aligned with the target shallow water behavior feature vector set in terms of feature distribution, a first feature similarity loss function is introduced to calculate the first similarity loss data; the enhancement parameter strategy set is re-obtained in the enhancement parameter action space according to the first similarity loss data until the first similarity loss data is minimized, thus obtaining the first type of enhancement processing parameters.

[0014] In a second aspect, this application provides a system for enhancing zebrafish behavior images. The system comprises: a zebrafish image set input module: inputting a zebrafish image set, extracting water layer-related features from each zebrafish image in the set, and dividing the zebrafish image set into a first-class zebrafish image set including a shallow-water label and a second-class zebrafish image set including a deep-water label according to the water layer-related features; a dual-branch enhancement adaptation model training module: training a dual-branch enhancement adaptation model based on a pre-constructed first-class representative behavior feature sample library and a second-class representative behavior feature sample library, wherein the dual-branch enhancement adaptation model includes a shared feature extraction network, a dual-branch feature multi-layer convolutional network, and a dual-branch enhancement parameter adaptation network; and a zebrafish behavior enhanced image set acquisition module: analyzing the first-class and second-class zebrafish image sets using the dual-branch enhancement adaptation model, and processing the first-class and second-class zebrafish image sets according to the obtained first-class and second-class enhancement processing parameters to obtain a zebrafish behavior enhanced image set.

[0015] In summary, one or more technical solutions provided in this application achieve differentiated enhancement of shallow-water zebrafish images and deep-water zebrafish images, adapt to the scene characteristics of different water layers, improve the accuracy of behavioral feature extraction in shallow and deep water scenes, ensure the integrity of behavioral features while improving image quality, enhance the ability to capture multi-scale behavioral details and the recognition of behavioral features, and ensure the accuracy of zebrafish behavior analysis. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application 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 merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the method for enhancing zebrafish behavior images is provided for this application.

[0018] Figure 2 This application provides a schematic diagram of the structure of an enhancement processing system for zebrafish behavior images.

[0019] Figure labeling: Zebrafish image set input module M100, dual-branch enhancement adaptation model training module M200, zebrafish behavior enhancement image set acquisition module M300. Detailed Implementation

[0020] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0021] Example 1: The present application will be described in detail below with reference to the accompanying drawings, as follows... Figure 1 As shown, this application provides a method for enhancing zebrafish behavior images, wherein the method includes:

[0022] S1: Input a set of zebrafish images, extract water layer-related features from each zebrafish image in the set, and divide the set of zebrafish images into a first-class set of zebrafish images with shallow water labels and a second-class set of zebrafish images with deep water labels according to the water layer-related features; S2: Train a bi-branch enhancement adaptation model based on a pre-constructed first-class representative behavioral feature sample library and a second-class representative behavioral feature sample library. The bi-branch enhancement adaptation model includes a shared feature extraction network, a bi-branch feature multi-layer convolutional network, and a bi-branch enhancement parameter adaptation network.

[0023] Specifically, the zebrafish image set refers to a dataset containing multiple images of zebrafish behavior. Each image records the behavior of zebrafish in a specific water layer, including swimming and foraging. Zebrafish images may have different lighting, contrast, and detail features due to differences in water layer. Water layer-related features refer to feature parameters related to the water layer environment in which the zebrafish image is located, used to distinguish between shallow and deep water scenes. These include global average brightness, background texture variance, and histogram distribution, which can reflect the influence of different water layers on the image. Shallow water labels and deep water labels are used to identify the type of water layer to which the image belongs. Furthermore, a shallow water label indicates that the image was collected in a shallow water area with strong lighting and a simple background, while a deep water label indicates that the image was collected in a deep water area with weak lighting and a complex background.

[0024] The dual-branch enhancement adaptation model is a deep learning model for processing zebrafish images at different water depths. It includes a shared feature extraction network, a dual-branch feature multi-layer convolutional network, and a dual-branch enhancement parameter adaptation network. The dual-branch enhancement adaptation model can simultaneously process shallow and deep water layers, generating the most suitable enhancement parameters for each layer. Furthermore, the shared feature extraction network is used to extract general features of zebrafish images, which exist in both shallow and deep water images, including the zebrafish's outline and movement trajectory. By sharing these features, the behavioral characteristics of zebrafish can be better identified and processed. The dual-branch feature multi-layer convolutional network contains two branches, one for shallow water images and one for deep water images. Each branch extracts detailed features of specific water layer images through multi-layer convolution operations, such as high-brightness details in shallow water images and low-contrast details in deep water images. The dual-branch enhancement parameter adaptation network generates enhancement parameters suitable for both shallow and deep water images based on the extracted features. Enhancement parameters include brightness adjustment, contrast enhancement, and sharpening, which are used to improve image quality and highlight behavioral characteristics.

[0025] Execution Steps: First, input a set of zebrafish images. Divide the images into shallow and deep water categories by extracting water layer-related features. These features include global average brightness and background texture variance. Specifically, based on a pre-built training dataset, a support vector machine classifier is used. The classification accuracy must exceed a preset threshold to effectively distinguish images from different water layers, ensuring classification reliability. Next, train a two-branch enhancement adaptation model using a pre-built sample library of representative behavioral features from shallow and deep water. Specifically, the two-branch enhancement adaptation model extracts general behavioral features from the images through a shared feature extraction network, and extracts specific behavioral features from shallow and deep water images separately through a two-branch feature multi-layer convolutional network. The two-branch enhancement parameter adaptation network generates enhancement parameters suitable for shallow and deep water images based on these features. For example, for shallow water images, the model may generate enhancement parameters to improve image contrast and sharpening intensity; for deep water images, it may adjust brightness and noise suppression intensity to highlight behavioral features. Preferably, this differentiated enhancement significantly improves image quality while preserving the integrity of behavioral features to the maximum extent, providing a high-quality data foundation for behavioral analysis.

[0026] S3: The first type of zebrafish image set and the second type of zebrafish image set are analyzed using the dual-branch enhancement adaptation model. The first type of enhancement processing parameters and the second type of enhancement processing parameters are processed accordingly to obtain the zebrafish behavior enhancement image set.

[0027] Specifically, the first type of enhancement parameters refers to the enhancement parameters generated for the shallow-water zebrafish image set. These parameters are obtained based on the feature analysis results of the shallow-water branch in the bi-branch enhancement adaptation model and are used to optimize the visual effect of the shallow-water images, including adjusting brightness, contrast, and sharpening intensity to highlight behavioral characteristics in shallow-water scenes. The second type of enhancement parameters refers to the enhancement parameters generated for the deep-water zebrafish image set. These parameters are obtained based on the feature analysis results of the deep-water branch in the bi-branch enhancement adaptation model and are used to optimize the visual effect of the deep-water images, including increasing brightness, reducing noise, and enhancing contrast to highlight behavioral characteristics in deep-water scenes. The zebrafish behavior enhancement image set refers to the set of zebrafish images after enhancement processing. These zebrafish behavior enhancement images, after optimization, can more clearly display the behavioral characteristics of zebrafish, including swimming trajectories and posture changes.

[0028] Execution steps: The dual-branch enhancement adaptation model analyzes the classified first-class zebrafish image set and the second-class zebrafish image set. It extracts general behavioral features of the images through a shared feature extraction network, and inputs these features into the shallow water branch and deep water branch of the dual-branch feature multilayer convolutional network. The shallow water branch and deep water branch extract specific behavioral features of the shallow water and deep water images respectively, and the corresponding enhancement processing parameters are generated through the dual-branch enhancement parameter adaptation network. Using the first-class and second-class enhancement processing parameters, the visual effect of the images is optimized, significantly improving image quality. Preferably, after enhancement, the high-brightness areas of the background in the shallow water images are clearer, and the outline and movement trajectory of the zebrafish are more prominent; after enhancement, the background noise in the deep water images is effectively suppressed, low-contrast details are enhanced, and the behavioral characteristics of the zebrafish are more obvious. Through this differentiated enhancement processing, the resulting zebrafish behavior-enhanced image set can more accurately reflect the behavioral characteristics of zebrafish, thus significantly improving the reliability of zebrafish behavior analysis.

[0029] Furthermore, the method of this application includes dividing the zebrafish image set according to the water layer-related features:

[0030] Water-related features are extracted from each zebrafish image. These features include at least global average brightness features, histogram distribution features, background local binary variance, background texture variance features, and background edge distribution features. A training dataset is constructed, comprising a first-class zebrafish training image set with known shallow-water labels and a second-class zebrafish training image set with known deep-water labels, along with corresponding first-group and second-group water-related feature training samples. A support vector machine (SVM) is used as a base classifier to train a classification model on the training dataset, resulting in an SVM model with a classification accuracy greater than a preset accuracy. The converged SVM model performs label classification of the zebrafish image set based on the input water-related features.

[0031] Specifically, the global average brightness feature refers to the average brightness value of all pixels in an image, reflecting the overall illumination intensity of the image. In zebrafish images, shallow water images typically have a higher global average brightness, while deep water images have a lower one. The histogram distribution feature describes the distribution of pixel brightness values ​​in an image. By analyzing the histogram, we can understand information such as the image's contrast and brightness range. Furthermore, the histogram of shallow water images may be concentrated in higher brightness areas, while the histogram of deep water images is more biased towards lower brightness areas. The local binary variance of the background refers to the variance value calculated after binarizing the background region of the image. It is used to measure the texture complexity of the background region. A higher variance value indicates a more complex background texture, which is common in deep water images.

[0032] Background texture variance features refer to variance analysis of the texture of the background region of an image to reflect the uniformity of the background texture. Furthermore, due to insufficient lighting and water scattering, the background texture variance of deep-water images is usually large. Background edge distribution features refer to the analysis of edge distribution density and direction by detecting edge information in the background region. In deep-water images, the edge distribution may be sparse and the direction may not be obvious, while in shallow-water images, the opposite is true. The training dataset contains a set of zebrafish images labeled with water layer labels and their corresponding water layer-related feature samples, used to train the classification model so that it can accurately distinguish between shallow and deep-water images based on water layer features. Support vector machines are used to classify zebrafish images based on water layer-related features, separating data of different categories by finding the optimal hyperplane. The preset accuracy is used to measure whether the classification effect of the support vector machine model has reached the expected standard. For example, the preset accuracy is set to 95%. Training is considered complete only when the classification accuracy of the support vector machine model reaches or exceeds the preset accuracy.

[0033] Execution steps: Extract water-layer-related features from the zebrafish image set, including global average brightness, histogram distribution, local binary variance of the background, background texture variance, and background edge distribution, which can comprehensively reflect the lighting conditions, background complexity, and texture information of the image; construct a training dataset, including a first-class zebrafish training image set with known shallow-water labels and a second-class zebrafish training image set with known deep-water labels, as well as corresponding water-layer-related feature samples. The training dataset is used to train a support vector machine (SVM) classifier; during the SVM model training process, the performance of the classification model is optimized by adjusting hyperparameters. The converged SVM model can accurately classify the zebrafish image set according to the input water-layer-related features, dividing the images into shallow-water and deep-water categories. Preferably, through accurate classification, it can be ensured that the shallow-water and deep-water images adopt the most suitable enhancement parameters respectively, providing an accurate classification basis for differential enhancement processing.

[0034] Furthermore, the method of this application includes training a classification model on the training dataset using a support vector machine as a base classifier.

[0035] A radial basis function kernel is introduced, which is obtained by configuring a 6-dimensional feature dimension, the feature importance weights corresponding to the 6-dimensional feature dimension, and the feature standard deviation; hyperparameters are initialized, including penalty parameters and kernel function parameters; the SMO optimization algorithm is used to train the model on the standardized training dataset to obtain Lagrange multipliers and biases, and the support vector machine trained based on the Lagrange multipliers and biases is used for model performance testing and optimization until a support vector machine model with a classification accuracy greater than the preset accuracy is obtained.

[0036] Specifically, the radial basis function (RBF) kernel is used in support vector machines to map input features to a high-dimensional space, thereby achieving non-linear classification. The form of the RBF kernel is: , where the first input vector This refers to a 6-dimensional feature vector extracted from the training dataset, the second input vector. This refers to another 6-dimensional feature vector extracted from the training dataset. This represents the output value of the radial basis function kernel, i.e., the first input vector. Second input vector Similarity measure between This represents the Euclidean distance between two vectors. It is the square of the Euclidean distance, used to control the decay of similarity. It is a kernel function parameter that controls the width of the mapping.

[0037] Indicates The exponential function of the exponent, which represents the square of the distance through... After scaling, it is mapped to the (0,1] interval. Specifically, when the first input vector... Second input vector When very close, A value close to 1 indicates that the two vectors are very similar in high-dimensional space; when the first input vector is close to 1, the two vectors are very similar in high-dimensional space. Second input vector When they are far apart, A value close to 0 indicates that the two vectors are not similar in the high-dimensional space; the 6-dimensional feature refers to the feature vector used to train the support vector machine model containing 6 dimensions, each corresponding to a water layer-related feature, including global average brightness, histogram distribution, and background local binary variance; the feature importance weight represents the importance of each feature in the classification task. In the radial basis kernel function, the influence of different features on the classification results is adjusted by assigning weights to each feature dimension.

[0038] The standard deviation of a feature measures the range of numerical distribution for each feature dimension. In kernel function configuration, the standard deviation is used to adjust the normalization degree of features, ensuring that the contributions of different features in high-dimensional space are relatively balanced. In support vector machine models, hyperparameters refer to parameters that need to be set manually, including penalty parameters. and kernel function parameters Penalty parameters The kernel function parameters control the degree of penalty for misclassification in the support vector machine model. The width of the feature map is controlled. The Sequential Minimal Optimization (SMO) algorithm is an efficient quadratic programming algorithm used to train support vector machine (SVM) models. SMO decomposes the original optimization problem into a series of smaller subproblems, thus efficiently calculating Lagrange multipliers and biases. Lagrange multipliers represent the contribution of each training sample to the classification hyperplane. By optimizing the Lagrange multipliers, the optimal classification hyperplane is found. In the SVM model, the bias term adjusts the position of the classification hyperplane to ensure the model better fits the training data.

[0039] Execution steps: A radial basis function kernel is introduced and configured according to the 6-dimensional feature dimension, feature importance weights, and feature standard deviations to map the input water layer-related features to a high-dimensional space in order to better capture the nonlinear relationships between features. Specifically, the 6-dimensional feature dimensions are global average brightness, histogram distribution, background local binary variance, background texture variance, background edge distribution, and image contrast. By assigning weights and standard deviations to each feature, the contribution of features in the high-dimensional space is adjusted, thereby improving the classification performance of the support vector machine model.

[0040] Initialization hyperparameters include penalty parameters and kernel function parameters The performance of the support vector machine (SVM) model is optimized by progressively adjusting the initial hyperparameters through grid search. The SMO optimization algorithm is used to train the model on the standardized training dataset. During training, the SMO algorithm efficiently calculates Lagrange multipliers and biases by decomposing the optimization problem, thereby determining the classification hyperplane. The SVM trained based on the Lagrange multipliers and biases is then tested and optimized until a SVM model with a classification accuracy greater than the preset accuracy is obtained. Preferably, by optimizing the kernel function and hyperparameters, the classification performance of the SVM model is improved, ensuring that the SVM model can accurately classify zebrafish images based on water layer-related features, providing a basis for differential enhancement processing, and thus maximizing the preservation of behavioral features while improving image quality.

[0041] Furthermore, the method of this application includes analyzing the first type of zebrafish image set and the second type of zebrafish image set using the dual-branch enhanced adaptation model:

[0042] The shared feature extraction network extracts first and second row feature vector sets from the first and second zebrafish image sets, respectively. The first row feature vector set is sent to the first branch of a dual-branch feature multi-layer convolutional network, and the second row feature vector set is sent to the second branch of the same network for multi-scale feature convolution, resulting in shallow water and deep water behavior feature vector sets. The dual-branch enhancement parameter adaptation network optimizes the shallow water behavior feature vector set using a defined first feature similarity loss function and optimizes the deep water behavior feature vector set using a defined second feature similarity loss function, resulting in first and second enhancement processing parameters. The enhancement parameter action space includes brightness, contrast, saturation, sharpening intensity, color balance, and noise suppression intensity parameters.

[0043] Specifically, the shared feature extraction network is used to extract general behavioral feature vectors from zebrafish images, containing behavioral information of zebrafish in the images, including swimming trajectory and posture. This information exists in both shallow and deep water images and can therefore be shared and extracted. The behavioral feature vector set refers to the set of feature vectors extracted from zebrafish images. The behavioral feature vectors contain the behavioral features of zebrafish in the images and are used for feature analysis and enhancement processing. The dual-branch feature multi-layer convolutional network contains two branches, which perform feature extraction for shallow and deep water images respectively. Each branch extracts multi-scale behavioral features of specific water layers through multi-layer convolution operations to capture detailed information at different scales.

[0044] The first branch and the second branch feature multi-layer convolutional network correspond to feature extraction branches for shallow and deep water images, respectively. Each branch extracts multi-scale behavioral features of a specific water layer image through multi-layer convolutional operations. The feature similarity loss function is used to measure the similarity between the enhanced image features and the target features. By minimizing this loss function, the dual-branch enhancement adaptation model can optimize the enhancement parameters, making the enhanced image features closer to the target features. The enhancement parameter action space refers to the range of enhancement parameters that the dual-branch enhancement adaptation model can adjust, including brightness parameters, contrast parameters, saturation parameters, sharpening intensity parameters, color balance parameters, and noise suppression intensity parameters, which are used to adjust the visual effect of the image to highlight behavioral features.

[0045] Execution steps: A common set of behavioral feature vectors is extracted from the first and second sets of zebrafish images using a shared feature extraction network. The shallow water behavioral feature vector set is sent to the first branch of the dual-branch feature multi-layer convolutional network, and the deep water behavioral feature vector set is sent to the second branch. Each branch extracts multi-scale behavioral features through multi-layer convolutional operations. Specifically, the first branch extracts detailed features of shallow water images through 3 layers of convolutional operations, and the second branch extracts detailed features of deep water images through the same 3 layers of convolutional operations. Each convolutional operation can use convolutional kernels of different sizes to capture features at different scales.

[0046] The dual-branch enhancement parameter adaptation network optimizes the shallow water behavior feature vector set and the deep water behavior feature vector set in the enhancement parameter action space by defining the first feature similarity loss function and the second feature similarity loss function, respectively. Furthermore, for shallow water images, the enhanced image features are made closer to the target features by adjusting the brightness parameter, contrast parameter, and sharpening intensity parameter; for deep water images, the brightness parameter, noise suppression intensity parameter, and contrast parameter are adjusted to highlight the behavior features.

[0047] The first and second types of enhancement parameters are obtained and used to perform differential enhancement processing on shallow and deep water images, thereby improving image quality and preserving the integrity of behavioral features. Specifically, after enhancement processing, the background details of shallow water images are clearer and the swimming trajectory of zebrafish is more obvious; noise in deep water images is effectively suppressed and low-contrast details are enhanced. After this differential enhancement processing, the accuracy of behavioral feature extraction is improved, significantly enhancing the accuracy and reliability of zebrafish behavior analysis.

[0048] Furthermore, the method of this application includes:

[0049] The shared feature extraction network is obtained through training by connecting three layers of neural networks. The three layers of neural networks include pixel feature extraction in the first interval, pixel feature extraction in the second interval, and pixel feature extraction in the third interval. The first interval is 52×52, the second interval is 26×26, and the third interval is 13×13.

[0050] Specifically, the shared feature extraction network is used to extract general behavioral feature vectors from zebrafish images, which can reflect the behavioral information of zebrafish, including swimming trajectory and posture. This behavioral information of zebrafish exists in both shallow and deep water images, so it can be shared and extracted. The 3-layer neural network refers to the shared feature extraction network being composed of 3 layers of neural networks, with each layer responsible for extracting pixel features at different scales. This hierarchical structure can capture multi-scale information in the image, thus reflecting the behavioral characteristics of zebrafish more comprehensively.

[0051] The first, second, and third intervals refer to the feature map size processed by each layer of the three-layer neural network. The first interval is 52×52, the second interval is 26×26, and the third interval is 13×13. As the number of network layers increases, the size of the feature map gradually decreases, but the level of abstraction of the features gradually increases. Pixel feature extraction refers to the process by which each layer of the neural network processes the input image or feature map to extract pixel-level features related to zebrafish behavior, including the zebrafish's outline, texture, and motion trajectory, which can provide basic information for behavior analysis.

[0052] Execution steps: The shared feature extraction network is trained through a 3-layer neural network. Each layer is responsible for extracting pixel features at different scales. Specifically, the first layer of the neural network receives the original zebrafish image and extracts low-level features from the image, including edges and textures. The feature map size corresponding to the first interval is 52×52, which can capture larger-scale information in the image. Furthermore, for a 256×256 zebrafish image, after the first convolution operation, the feature map size is reduced to 52×52. At this time, the extracted features are mainly concentrated on the overall outline of the zebrafish and large texture changes.

[0053] The second layer of the neural network further processes the output of the first layer, extracting more abstract features. The feature map size corresponding to the second interval is 26×26, which can capture information at a medium scale. Furthermore, through the second layer convolution operation, the feature map size is reduced from 52×52 to 26×26. At this time, the extracted features include the local motion trajectory and posture changes of the zebrafish. The third layer of the neural network processes the output of the second layer, extracting higher-level features. The feature map size corresponding to the third interval is 13×13, which can capture information at a smaller scale. Furthermore, through the third layer convolution operation, the feature map size is reduced from 26×26 to 13×13. At this time, the extracted features include the subtle movements and behavioral details of the zebrafish.

[0054] Preferably, through this hierarchical structure, the shared feature extraction network can extract zebrafish behavioral features from different scales, thereby providing comprehensive and rich feature information for differential enhancement processing. By comparing the effects of using only a single-layer feature extraction network and a three-layer feature extraction network, it was found that the three-layer network can capture zebrafish behavioral features more accurately and improve the accuracy of behavioral feature extraction. This multi-scale feature extraction method can significantly improve the model's adaptability and robustness to images at different water layers, providing a higher quality feature foundation for differential enhancement processing, and thus improving the accuracy and reliability of zebrafish behavior analysis.

[0055] Furthermore, the method of this application includes:

[0056] The first branch feature multilayer convolutional network is constructed from K convolutional layers consisting of K first convolutional kernel clustering groups, and the second branch feature multilayer convolutional network is constructed from K convolutional layers consisting of K second convolutional kernel clustering groups. Specifically, the K first convolutional kernel clustering groups are obtained by selecting the first feature pixel scale set belonging to the first class of representative behavioral feature sample library and performing K-value clustering on the first feature pixel scale set; the K second convolutional kernel clustering groups are obtained by selecting the second feature pixel scale set belonging to the second class of representative behavioral feature sample library and performing K-value clustering on the first feature pixel scale set.

[0057] Specifically, the first branch feature multilayer convolutional network is the convolutional network part of the two-branch enhancement adaptation model that processes shallow-water zebrafish images. It consists of multiple convolutional layers, each using a specific convolutional kernel to extract features from the shallow-water image. The second branch feature multilayer convolutional network is the convolutional network part of the two-branch enhancement adaptation model that processes deep-water zebrafish images. It also consists of multiple convolutional layers, each using a specific convolutional kernel to extract features from the deep-water image. The convolutional kernel clustering group refers to the set of convolutional kernels extracted from the feature sample library through a clustering algorithm. Each convolutional kernel clustering group contains multiple convolutional kernels used to capture features of images at a specific water layer. The first feature pixel scale set is the set of feature pixel scales selected from the feature sample library representing behavior at the shallow water layer, used to describe the feature distribution of the shallow-water image. The second feature pixel scale set is the set of feature pixel scales selected from the feature sample library representing behavior at the deep water layer, used to describe the feature distribution of the deep-water image. K-value clustering is used to divide the feature pixel scale set into K clusters, each cluster corresponding to a convolutional kernel. These convolutional kernels capture features of images at a specific water layer.

[0058] Execution steps: In the setup of the dual-branch feature multilayer convolutional network, the first branch feature multilayer convolutional network and the second branch feature multilayer convolutional network extract features from shallow water and deep water zebrafish images, respectively. Each branch consists of multiple convolutional layers, and each convolutional layer uses a specific convolutional kernel to extract image features.

[0059] The first branch feature multi-layer convolutional network corresponds to shallow water images. Specifically, it extracts a first feature pixel scale set from a first-class representative behavioral feature sample library. This first feature pixel scale set can describe the key features of the shallow water image, including high-brightness areas and clear edges. K-value clustering is performed on the first feature pixel scale set to obtain K first convolutional kernel cluster groups. Each cluster corresponds to one convolutional kernel. These convolutional kernels capture specific features of the shallow water image, and each kernel is used to extract different features of the shallow water image, including edge features and texture features. The first branch feature multi-layer convolutional network is constructed from K convolutional layers consisting of K first convolutional kernel cluster groups. Furthermore, using these K convolutional layers, each convolutional layer extracts multi-scale features of the shallow water image through convolution operations. For example, the first convolutional layer of the first branch feature multi-layer convolutional network extracts edge features of the shallow water image, the second convolutional layer extracts texture features, and the third convolutional layer extracts motion trajectory features.

[0060] The second-branch feature multi-layer convolutional network corresponds to deep-water images. Specifically, it extracts a second feature pixel scale set from a second-class representative behavioral feature sample library. This second feature pixel scale set can describe the key features of deep-water images, including low-brightness areas and complex background textures. K-value clustering is performed on the second feature pixel scale set to obtain K second convolutional kernel cluster groups. Each cluster corresponds to one convolutional kernel. These convolutional kernels capture specific features of deep-water images, and each kernel is used to extract different features of deep-water images, including low-contrast features and noise features. The second-branch feature multi-layer convolutional network is constructed from K convolutional layers consisting of K second convolutional kernel cluster groups. Furthermore, using these K convolutional layers, each convolutional layer extracts multi-scale features of the deep-water image through convolution operations. For example, the first convolutional layer of the second-branch feature multi-layer convolutional network extracts low-contrast features of the deep-water image, the second convolutional layer extracts background texture features, and the third convolutional layer extracts the contour features of zebrafish.

[0061] Preferably, a multi-layer convolutional network with first-branch features and a multi-layer convolutional network with second-branch features are used for targeted feature extraction, which significantly improves the adaptability to images of different water layers and provides a higher quality feature foundation for enhanced processing. Through this clustering-based convolutional kernel setting, the dual-branch feature multi-layer convolutional network can efficiently extract specific features for shallow and deep water images, thereby improving the accuracy and reliability of zebrafish behavior analysis.

[0062] Furthermore, by optimizing the shallow water behavior feature vector set in the enhanced parameter action space through a defined first feature similarity loss function, the method of this application includes:

[0063] An initial enhancement parameter strategy set is obtained in the enhancement parameter action space; the shallow water behavior feature vector set is enhanced according to the initial enhancement parameter strategy set, and the enhanced shallow water behavior feature vector set is aligned with the target shallow water behavior feature vector set in terms of feature distribution. A first feature similarity loss function is introduced to calculate the first similarity loss data; the enhancement parameter strategy set is re-obtained in the enhancement parameter action space according to the first similarity loss data until the first similarity loss data is minimized to obtain the first type of enhancement processing parameters.

[0064] Specifically, the enhancement parameter action space refers to the set of all enhancement parameters adjusted, including brightness, contrast, saturation, sharpening intensity, color balance, and noise suppression intensity parameters, used to adjust the visual effect of the image to highlight behavioral features. The initial enhancement parameter strategy set refers to a random or preset set of initial parameters in the enhancement parameter action space, used to initiate the enhancement process. The initial enhancement parameter strategy set is the starting point for parameter enhancement. Feature distribution alignment refers to matching the enhanced feature vector with the target feature vector in terms of distribution, making the enhanced image features as close as possible to the target features. This is usually achieved by calculating the similarity or distance between feature vectors. The first feature similarity loss function is a function used to measure the similarity between the enhanced shallow water behavior feature vector set and the target shallow water behavior feature vector set. By minimizing this loss function, the dual-branch enhancement parameter adaptation network can optimize the enhancement parameters, making the enhanced features closer to the target features. The first similarity loss data is a value calculated by the first feature similarity loss function, representing the degree of difference between the enhanced features and the target features. The dual-branch enhancement parameter adaptation network adjusts the enhancement parameters by optimizing the first similarity loss data.

[0065] Execution steps: Obtain an initial set of enhancement parameter strategies in the enhancement parameter action space. The initial set of enhancement parameter strategies is the starting point of the enhancement algorithm and is used to initiate the enhancement processing of the shallow water behavior feature vector set. Enhance the shallow water behavior feature vector set according to this initial set of enhancement parameter strategies. Align the enhanced feature vector set with the target shallow water behavior feature vector set. The target shallow water behavior feature vector set is extracted from high-quality shallow water images and has clear zebrafish outlines and high-contrast backgrounds.

[0066] To measure the similarity between the enhanced features and the target features, a first feature similarity loss function is introduced. This function calculates the similarity loss data between the enhanced feature vector set and the target feature vector set. For example, the first feature similarity loss function uses mean squared error to calculate the initial similarity loss data. Based on the first similarity loss data, a new set of enhancement parameter strategies is obtained in the enhancement parameter action space. The enhancement parameters are continuously adjusted in the enhancement parameter action space using optimization algorithms such as gradient descent until the first similarity loss data is minimized. The enhancement parameter strategy set currently obtained from the enhancement parameter action space is used as the first type of enhancement processing parameters. Preferably, the optimized enhancement processing generates the most suitable enhancement parameters for shallow water zebrafish images, improving the accuracy of behavioral feature extraction from shallow water images and significantly enhancing the accuracy and reliability of zebrafish behavior analysis.

[0067] In summary, the beneficial effects of the embodiments of this application are:

[0068] Because an input set of zebrafish images is used, water layer-related features of each zebrafish image in the set are extracted. Based on these features, the set is divided into a first-class set (including shallow-water labels) and a second-class set (including deep-water labels). A bi-branch enhancement adaptation model is trained using pre-constructed representative behavioral feature sample libraries for both classes. This model includes a shared feature extraction network, a bi-branch feature multi-layer convolutional network, and a bi-branch enhancement parameter adaptation network. The bi-branch enhancement adaptation model is then used to analyze both the first and second-class zebrafish image sets. The first and second-class enhancement processing parameters are then used to process these images, resulting in a zebrafish behavior-enhanced image set. This application provides a method and system for enhancing zebrafish behavior images, achieving differentiated enhancement of shallow-water and deep-water zebrafish images, adapting to the scene characteristics of different water layers, improving the accuracy of behavioral feature extraction in shallow and deep water scenes, ensuring the integrity of behavioral features while improving image quality, enhancing the ability to capture multi-scale behavioral details and the recognition of behavioral features, and ensuring the accuracy of zebrafish behavior analysis.

[0069] Example 2, based on the same inventive concept as the zebrafish behavior image enhancement processing method in the aforementioned examples, such as... Figure 2 As shown in the embodiment of this application, an enhancement processing system for zebrafish behavior images is provided, wherein the system includes:

[0070] Zebrafish image set input module M100: Inputs a zebrafish image set, extracts water layer related features of each zebrafish image in the zebrafish image set, and divides the zebrafish image set into a first type of zebrafish image set including shallow water labels and a second type of zebrafish image set including deep water labels according to the water layer related features.

[0071] The dual-branch augmentation adaptation model training module M200 trains a dual-branch augmentation adaptation model based on a pre-constructed first-class representative behavior feature sample library and a second-class representative behavior feature sample library. The dual-branch augmentation adaptation model includes a shared feature extraction network, a dual-branch feature multi-layer convolutional network, and a dual-branch augmentation parameter adaptation network.

[0072] Zebrafish behavior enhancement image set acquisition module M300: wherein, the first type of zebrafish image set and the second type of zebrafish image set are analyzed using the dual-branch enhancement adaptation model, and the first type of zebrafish image set and the second type of zebrafish image set are processed according to the obtained first type of enhancement processing parameters and second type of enhancement processing parameters to obtain the zebrafish behavior enhancement image set.

[0073] Furthermore, the zebrafish image set input module M100 is used to perform the following method:

[0074] Water-related features are extracted from each zebrafish image. These features include at least global average brightness features, histogram distribution features, background local binary variance, background texture variance features, and background edge distribution features. A training dataset is constructed, comprising a first-class zebrafish training image set with known shallow-water labels and a second-class zebrafish training image set with known deep-water labels, along with corresponding first-group and second-group water-related feature training samples. A support vector machine (SVM) is used as a base classifier to train a classification model on the training dataset, resulting in an SVM model with a classification accuracy greater than a preset accuracy. The converged SVM model performs label classification of the zebrafish image set based on the input water-related features.

[0075] Furthermore, the zebrafish image set input module M100 is also used to perform the following method:

[0076] A radial basis function kernel is introduced, which is obtained by configuring a 6-dimensional feature dimension, the feature importance weights corresponding to the 6-dimensional feature dimension, and the feature standard deviation; hyperparameters are initialized, including penalty parameters and kernel function parameters; the SMO optimization algorithm is used to train the model on the standardized training dataset to obtain Lagrange multipliers and biases, and the support vector machine trained based on the Lagrange multipliers and biases is used for model performance testing and optimization until a support vector machine model with a classification accuracy greater than the preset accuracy is obtained.

[0077] Furthermore, the zebrafish behavior enhancement image set acquisition module M300 is used to perform the following method:

[0078] The shared feature extraction network extracts first and second row feature vector sets from the first and second zebrafish image sets, respectively. The first row feature vector set is sent to the first branch of a dual-branch feature multi-layer convolutional network, and the second row feature vector set is sent to the second branch of the same network for multi-scale feature convolution, resulting in shallow water and deep water behavior feature vector sets. The dual-branch enhancement parameter adaptation network optimizes the shallow water behavior feature vector set using a defined first feature similarity loss function and optimizes the deep water behavior feature vector set using a defined second feature similarity loss function, resulting in first and second enhancement processing parameters. The enhancement parameter action space includes brightness, contrast, saturation, sharpening intensity, color balance, and noise suppression intensity parameters.

[0079] Furthermore, the zebrafish behavior enhancement image set acquisition module M300 is also used to perform the following method:

[0080] The shared feature extraction network is obtained through training by connecting three layers of neural networks. The three layers of neural networks include pixel feature extraction in the first interval, pixel feature extraction in the second interval, and pixel feature extraction in the third interval. The first interval is 52×52, the second interval is 26×26, and the third interval is 13×13.

[0081] Furthermore, the zebrafish behavior enhancement image set acquisition module M300 is also used to perform the following method:

[0082] The first branch feature multilayer convolutional network is constructed from K convolutional layers consisting of K first convolutional kernel clustering groups, and the second branch feature multilayer convolutional network is constructed from K convolutional layers consisting of K second convolutional kernel clustering groups. Specifically, the K first convolutional kernel clustering groups are obtained by selecting the first feature pixel scale set belonging to the first class of representative behavioral feature sample library and performing K-value clustering on the first feature pixel scale set; the K second convolutional kernel clustering groups are obtained by selecting the second feature pixel scale set belonging to the second class of representative behavioral feature sample library and performing K-value clustering on the first feature pixel scale set.

[0083] Furthermore, the zebrafish behavior enhancement image set acquisition module M300 is also used to perform the following method:

[0084] An initial enhancement parameter strategy set is obtained in the enhancement parameter action space; the shallow water behavior feature vector set is enhanced according to the initial enhancement parameter strategy set, and the enhanced shallow water behavior feature vector set is aligned with the target shallow water behavior feature vector set in terms of feature distribution. A first feature similarity loss function is introduced to calculate the first similarity loss data; the enhancement parameter strategy set is re-obtained in the enhancement parameter action space according to the first similarity loss data until the first similarity loss data is minimized to obtain the first type of enhancement processing parameters.

[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Figure 1 The zebrafish behavior image enhancement processing method and specific examples in Example 1 are also applicable to the zebrafish behavior image enhancement processing system of this embodiment. Through the foregoing detailed description of the zebrafish behavior image enhancement processing method, those skilled in the art can clearly understand the zebrafish behavior image enhancement processing system of this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0086] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. 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 this application. Therefore, this application 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.

[0087] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for enhancing zebrafish behavioral images, characterized in that, The method includes: Input a zebrafish image set, extract water layer related features of each zebrafish image in the zebrafish image set, and divide the zebrafish image set into a first type of zebrafish image set including shallow water labels and a second type of zebrafish image set including deep water labels according to the water layer related features. A dual-branch enhancement adaptation model is trained based on a pre-constructed first-class representative behavior feature sample library and a second-class representative behavior feature sample library. The dual-branch enhancement adaptation model includes a shared feature extraction network, a dual-branch feature multi-layer convolutional network, and a dual-branch enhancement parameter adaptation network. Specifically, the dual-branch enhancement adaptation model is used to analyze the first type of zebrafish image set and the second type of zebrafish image set. Based on the obtained first type of enhancement processing parameters and second type of enhancement processing parameters, the first type of zebrafish image set and the second type of zebrafish image set are processed accordingly to obtain a zebrafish behavior enhancement image set. The method for analyzing the first type of zebrafish image set and the second type of zebrafish image set using the dual-branch enhanced adaptation model includes: The shared feature extraction network extracts a first set of behavioral feature vectors and a second set of behavioral feature vectors from the first set of zebrafish images and the second set of zebrafish images. The shared feature extraction network is used to extract general behavioral feature vectors from zebrafish images, which can reflect the behavioral information of zebrafish, including swimming trajectory and posture. The first set of behavioral feature vectors is sent to the first branch of the dual-branch feature multi-layer convolutional network, and the second set of behavioral feature vectors is sent to the second branch of the dual-branch feature multi-layer convolutional network for multi-scale feature convolution to obtain the shallow water behavioral feature vector set and the deep water behavioral feature vector set. The dual-branch enhancement parameter adaptation network optimizes the shallow water behavior feature vector set in the enhancement parameter action space through a defined first feature similarity loss function, and optimizes the deep water behavior feature vector set in the enhancement parameter action space through a defined second feature similarity loss function, thereby obtaining the first type of enhancement processing parameters and the second type of enhancement processing parameters. The first branch feature multilayer convolutional network is constructed from K convolutional layers consisting of K first convolutional kernel clusters, and the second branch feature multilayer convolutional network is constructed from K convolutional layers consisting of K second convolutional kernel clusters. K first convolutional kernel clusters are obtained by selecting the first feature pixel scale set of the first class representative behavioral feature sample library and performing K-value clustering on the first feature pixel scale set; K second convolutional kernel clusters are obtained by selecting the second feature pixel scale set of the second class representative behavioral feature sample library and performing K-value clustering on the first feature pixel scale set; The enhancement parameter action space includes brightness parameter, contrast parameter, saturation parameter, sharpening intensity parameter, color balance parameter, and noise suppression intensity parameter; The optimization of the shallow water behavior feature vector set in the enhanced parameter action space is performed using a defined first feature similarity loss function, and the method includes: Obtain the initial enhancement parameter strategy group in the enhancement parameter action space; The shallow water behavior feature vector set is enhanced according to the initial enhancement parameter strategy group. The enhanced shallow water behavior feature vector set is aligned with the target shallow water behavior feature vector set in terms of feature distribution. The first feature similarity loss function is introduced to calculate the first similarity loss data. Based on the first similarity loss data, the enhancement parameter strategy group is re-acquired in the enhancement parameter action space until the first similarity loss data is minimized, thus obtaining the first type of enhancement processing parameters.

2. The method as described in claim 1, characterized in that, The method for dividing the zebrafish image set according to the water layer-related features includes: Extract water layer related features from each zebrafish image. The water layer related features include at least global average brightness features, histogram distribution features, background local binary variance, background texture variance features, and background edge distribution features. Construct a training dataset, which includes a training image set of the first type of zebrafish with known shallow water labels and a training image set of the second type of zebrafish with known deep water labels, as well as corresponding first set of water layer-related feature training samples and second set of water layer-related feature training samples. A support vector machine model is used as a base classifier to train a classification model on the training dataset, resulting in a support vector machine model with a classification accuracy greater than a preset accuracy. The convergent support vector machine model performs label classification of the zebrafish image set based on the input water layer-related features.

3. The method as described in claim 2, characterized in that, The method for training a classification model on the training dataset using a support vector machine as a base classifier includes: A radial basis function kernel is introduced, which is obtained by configuring a 6-dimensional feature dimension, the feature importance weights corresponding to the 6-dimensional feature dimension, and the feature standard deviation; Initialize the hyperparameters, which include penalty parameters and kernel function parameters; The SMO optimization algorithm is used to train the model on the standardized training dataset to obtain Lagrange multipliers and biases. The support vector machine trained based on the Lagrange multipliers and biases is then used for model performance testing and optimization until a support vector machine model with a classification accuracy greater than the preset accuracy is obtained.

4. The method as described in claim 1, characterized in that, The shared feature extraction network is obtained through training by connecting three layers of neural networks. The three layers of neural networks include pixel feature extraction in the first interval, pixel feature extraction in the second interval, and pixel feature extraction in the third interval. The feature map size corresponding to the first interval is 52×52, the feature map size corresponding to the second interval is 26×26, and the feature map size corresponding to the third interval is 13×13.

5. A system for enhancing zebrafish behavioral images, characterized in that, The system comprises the following steps for implementing the zebrafish behavior image enhancement processing method according to any one of claims 1-4: Zebrafish image set input module: Input zebrafish image set, extract water layer related features of each zebrafish image in the zebrafish image set, and divide the zebrafish image set into a first type of zebrafish image set including shallow water label and a second type of zebrafish image set including deep water label according to the water layer related features; Dual-branch augmentation adaptation model training module: Train a dual-branch augmentation adaptation model based on a pre-constructed first-class representative behavior feature sample library and a second-class representative behavior feature sample library. The dual-branch augmentation adaptation model includes a shared feature extraction network, a dual-branch feature multi-layer convolutional network, and a dual-branch augmentation parameter adaptation network. The zebrafish behavior enhancement image set acquisition module: wherein, the first type of zebrafish image set and the second type of zebrafish image set are analyzed using the dual-branch enhancement adaptation model, and the first type of zebrafish image set and the second type of zebrafish image set are processed according to the obtained first type of enhancement processing parameters and second type of enhancement processing parameters to obtain the zebrafish behavior enhancement image set.

Images