Image enhancement processing method for fishery resource statistics

Abstract

The invention provides an image enhancement processing method for fishery resource statistics, belongs to the field of image enhancement, and aims to construct a color prior feature enhancement network which comprises a high-frequency information enhancement module and a low-frequency information enhancement module. The high-frequency information enhancement module strengthens detail features in the fishery resource statistical image, the low-frequency information enhancement module focuses on accurate correction of image colors, the overall content and color expression of the fishery resource statistical image are better reserved, the frequency information interaction module is constructed, interaction between high-frequency information and low-frequency information is achieved, and the accuracy of the fishery resource statistical image is improved. And color information in the low-frequency information and texture and detail features in the high-frequency information are effectively fused, so that fishery resource statistical image enhancement is realized.

Description

Technical Field

[0001] The invention belongs to the field of image enhancement, and in particular relates to an image enhancement processing method for fishery resource statistics. Background Art

[0002] In fishery resource management, it is crucial to accurately obtain information on the number, distribution and behavioral characteristics of fish populations. Due to the complexity of the underwater environment, the quality of fishery resource statistical images is greatly affected, resulting in low contrast, high noise and color distortion in underwater images, which in turn affects the accuracy of fish identification and statistics. Therefore, how to effectively enhance the images collected by underwater cameras has become a key technology to improve the accuracy of fishery resource statistics. Image enhancement technology mainly includes denoising, contrast enhancement, color correction, and edge detection. By removing the scattered light effect and lighting unevenness in the underwater environment, it can restore the details and color information of the image and enhance the visibility of fish. In this way, the accuracy of subsequent fish identification, classification and quantity statistics will be greatly improved, providing a reliable data basis for the accurate assessment of fishery resources.

[0003] In recent years, the excellent performance of diffusion-based methods in image restoration tasks has attracted widespread attention. Although the standard diffusion model shows sufficient ability, unforeseen artifacts may occur due to the introduction of diversity in the sampling process from randomly generated Gaussian noise to the image. In addition, the diffusion model has limited ability to focus on fine-grained information, resulting in the neglect of texture and details. Most of the previous methods are based on the original pixel space of the image, and the exploration of the frequency space characteristics of underwater images is limited, which makes it impossible to effectively utilize the representation ability of the deep model to generate high-quality images. In order to solve the above problems, the present invention proposes an image enhancement processing method for fishery resource statistics, constructs a feature information enhancement module to fully explore the frequency domain information in the fishery resource statistical image, realizes detail enhancement of high-frequency information and color correction of low-frequency information, and at the same time, realizes the interaction of high-frequency information and low-frequency information, further enhancing the image quality. Summary of the invention

[0004] The present invention provides an image enhancement processing method for fishery resource statistics, which aims to construct a color prior feature enhancement network, use discrete wavelet transform to divide an image into high-frequency information containing global structure and texture details and low-frequency information containing content and color, enhance the frequency information in wavelet space, separate and process the high-frequency information and the low-frequency information, respectively realize detail enhancement and color correction, and construct a frequency information interaction module, through which the interaction of high-frequency information and low-frequency information is realized, thereby realizing image enhancement.

[0005] The present invention aims to propose a color prior feature enhancement network and provide an image enhancement processing method for fishery resource statistics, which includes the following steps.

[0006] S1. Collection of statistical images of fishery resources. Install the underwater camera on an underwater robot, cruise and shoot in specific waters, obtain underwater statistical images of fishery resources, add labels to the obtained statistical images of fishery resources and make them into a data set.

[0007] S2. Construct a color prior feature enhancement network, which consists of N feature information enhancement modules. The feature information enhancement module includes a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module.

[0008] S3. Construct a low-frequency information enhancement module, including SE attention mechanism, regularization, GELU activation function, fully connected layer and multi-scale color information extraction module, to realize the color correction of low-frequency information of fishery resource statistical images.

[0009] S4. Construct a high-frequency information enhancement module, including a 1×1 convolutional layer, a deep convolutional layer, an attention mechanism, and a multi-layer perceptron, to enhance the high-frequency information details of fishery resource statistical images.

[0010] S5. Build a frequency information interaction module, including a 1×1 convolution layer, a 3×3 deep convolution layer, and a softmax activation function, to achieve mutual reinforcement of high- and low-frequency information.

[0011] S6. Processing of fishery resource statistical images: inputting the fishery resource statistical images to be processed into a color prior feature enhancement network to obtain enhanced fishery resource statistical images.

[0012] Preferably, in step S2, for the color prior feature enhancement network, the fishery resource statistical image is input , H, W and C represent The height, width and channels of Decompose into high-frequency information and low-frequency information, , represents discrete wavelet transform, Represents the low-frequency information of the input fishery resources statistical image, including the content and color information of the input fishery resources statistical image. It represents the high-frequency information of the input fishery resource statistical image, including the detailed information of the global structure and texture. Then the high-frequency information and low-frequency information are input into the feature information enhancement module. The feature information enhancement module consists of a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module. After the low-frequency information enhancement module, , high frequency information After the high-frequency information enhancement module, , and Input to the frequency information interaction module to realize the information interaction between high-frequency information and low-frequency information. , represents the frequency information interaction module, Represents the low-frequency information after interaction, Represents the high-frequency information after interaction, and then and Add the elements together to get , represents the addition of elements, and Add the elements together to get , represents the addition of elements, and As the input of the next feature information enhancement module, after N feature information enhancement modules, and , and The final image is generated by inverse discrete wavelet transform , Represents inverse discrete wavelet transform.

[0013] Preferably, in step S2, the feature information enhancement module performs targeted enhancement on the low-frequency information and the high-frequency information respectively, and realizes the interaction between the high-frequency information and the low-frequency information through the frequency information interaction module, and independently processes the information of different frequency bands to improve the overall visual quality of the image. At the same time, through multiple frequency interactions and feature enhancements, the image features are gradually optimized, and the optimized high-frequency information and low-frequency information are reconstructed into an enhanced image in combination with the inverse discrete wavelet transform, which effectively maintains the global structural consistency and detail clarity of the image, and realizes detail enhancement and color correction.

[0014] Preferably, in step S3, for the low-frequency information enhancement module, the low-frequency information of the fishery resources statistical image is input , H, W and C represent height, width and channel respectively. First, the SE attention mechanism adaptively assigns weights to different feature channels, enhances important features, suppresses unimportant features, and uses Dropout for regularization to prevent overfitting of model training. The 3×3 convolutional layer and GELU activation function work together to capture low-frequency information while improving the representation ability of the model, and obtain the optimized low-frequency information. , represents the SE attention mechanism, represents Dropout regularization, represents a 3×3 convolutional layer, Represents the GELU activation function, and then uses the multi-scale color information extraction module to obtain more comprehensive color information , Represents the multi-scale color information extraction module, and finally Input to the fully connected layer to obtain enhanced low-frequency information .

[0015] Preferably, in step S3, the design of the low-frequency information enhancement module gradually optimizes the low-frequency information of the fishery resource statistical image through multiple steps. First, the SE attention mechanism is used to adaptively weight the feature channels, which can enhance important features and suppress irrelevant features, ensuring that the model focuses on information that plays a key role in fishery resource statistics. Dropout is used for regularization to effectively prevent the model from overfitting during training. Combined with the 3×3 convolutional layer and the GELU activation function, it can not only capture local features in the low-frequency information, but also improve the model's nonlinear representation ability and detail capture ability. The multi-scale color information extraction module further obtains richer color information from different scales, thereby providing the model with more comprehensive and accurate color features. Finally, the optimized low-frequency information is output through the fully connected layer, ensuring that the low-frequency information not only retains the overall content of the image, but also is further enhanced in color processing.

[0016] Preferably, in step S3, for the multi-scale color information extraction module, in order to process and extract the information of each color channel separately, firstly, the input fishery resources statistical image features are Extract the R, G, and B color channels respectively. , It represents the separation of color channels, uses 3×3 convolutional layers and ReLU activation functions to extract local features in color channels, and then uses the SE attention mechanism to assign different weights to the R, G, and B color channels to enhance important features and suppress unimportant features. The color features processed by the channel attention mechanism are fused by element-wise addition, and finally the fused color features are layer-normalized to obtain multi-scale color information. , represents a 3×3 convolutional layer, represents the ReLU activation function, represents the SE attention mechanism, Representation layer normalization processing.

[0017] Preferably, in step S3, the multi-scale color information extraction module realizes independent processing of each color channel information by extracting R, G, and B color channels respectively. This separate processing method enables the model to deeply explore the local features of each channel, enhance the feature expression through 3×3 convolutional layers and ReLU activation functions, introduce SE attention mechanism, assign weights to different color channels, effectively improve the attention of important features and suppress unimportant features, thereby optimizing the efficiency and accuracy of information extraction. Finally, the features of each channel are fused by element addition and layer normalization is performed to ensure the balance and consistency of color information. This design not only improves the representation ability of color features, but also enhances the adaptability and performance of the model in complex image processing, making it more effective and reliable in the analysis of fishery resource statistical images.

[0018] Preferably, in step S4, for the high-frequency information enhancement module, firstly, the high-frequency information of the fishery resources statistical image is Perform layer normalization , H, W and C represent height, width and channel respectively, The representation layer is normalized, and the processed features are obtained by passing through a 1×1 convolution layer and a 7×7 deep convolution layer, a 1×1 convolution layer and a 5×5 deep convolution layer, and a 1×1 convolution layer and a 3×3 deep convolution layer. , , , Represents query, key and value respectively, represents a 1×1 convolutional layer, represents a 7×7 depth convolutional layer, represents a 5×5 depth convolutional layer, represents a 3×3 deep convolutional layer, and then performs self-attention calculation, layer normalization and multi-layer perception operations on Q, K and V, calculates the weighted value of the feature, improves the model's attention to important features, and finally introduces residual connections to improve the stability of model training. , represents the self-attention mechanism, Representation layer normalization processing, represents a multi-layer perception operation, Represents addition of elements.

[0019] Preferably, in step S4, the high-frequency information enhancement module effectively improves the processing capability of high-frequency information of fishery resource statistical images. Through 1×1 convolutional layers and deep convolutional layers with different convolutional kernels, the module can capture multi-scale detail features from different receptive fields, and generate query, key and value features respectively. Then, the self-attention mechanism can focus on important details in high-frequency information by calculating the correlation between Q, K and V, and improve the model's recognition capability of key features. It is combined with a multi-layer perceptron to further process feature information. Finally, the residual connection is used to add the input high-frequency information to the processed features, which effectively avoids the gradient vanishing problem and maintains the continuous flow of information. This not only enhances the detail expression of the image, but also improves the training stability of the model.

[0020] Preferably, in step S5, the frequency information interaction module is used to promote the information interaction between low-frequency information and high-frequency information, realize the mutual reinforcement of high-frequency and low-frequency information, and input high-frequency information. and low frequency information , H, W and C represent height, width and channel respectively, high frequency information After a 1×1 convolutional layer, we get , represents a 1×1 convolutional layer, express Query, high frequency information After 1×1 convolution layer and 3×3 depth convolution layer, we get , represents a 1×1 convolutional layer, represents a 3×3 depth convolutional layer, express The value of low-frequency information After a 1×1 convolutional layer, we get , represents a 1×1 convolutional layer, express Key, low frequency information After 1×1 convolution layer and 3×3 depth convolution layer, we get , represents a 1×1 convolutional layer, represents a 3×3 depth convolutional layer, express The value of and Perform softmax normalization to get , Represents the softmax activation function, and finally and Multiply elements together to get high-frequency information after interaction , Represents element-wise multiplication, and Multiply elements together to get the low-frequency information after interaction , Represents element-wise multiplication.

[0021] Preferably, in step S5, the frequency information interaction module realizes mutual reinforcement of high-frequency and low-frequency information by promoting the interaction between low-frequency information and high-frequency information. The module first extracts the query of high-frequency information and the key of low-frequency information through a 1×1 convolution layer, and obtains their respective values ​​using a deep convolution layer. Then, the query and the key are processed through a softmax normalization operation to generate a weight matrix, which captures the relationship between high-frequency and low-frequency information. Finally, the weight matrix is ​​multiplied by the high-frequency and low-frequency values ​​to achieve targeted reinforcement of information. This design not only improves the accuracy and effectiveness of feature extraction, but also enhances the adaptability of the model when processing complex images, ensuring the combination of high-frequency details and low-frequency information.

[0022] Compared with the prior art, the present invention has the following beneficial effects: The color prior feature enhancement network provided by the present invention uses discrete wavelet transform to divide an image into high-frequency information containing global structure and texture details and low-frequency information containing content and color, enhances the frequency information in wavelet space, separates the high-frequency information and the low-frequency information, respectively handles the detail enhancement and color correction problems, and constructs a frequency information interaction module at the same time. Through this module, the interaction between the high-frequency information and the low-frequency information is realized, thereby enhancing the image quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of an image enhancement processing method for fishery resource statistics provided by the present invention; Figure 2 It is a color prior feature enhancement network structure diagram provided by the present invention; Figure 3 It is a structural diagram of a low-frequency information enhancement module provided by the present invention; Figure 4 It is a structural diagram of the high-frequency information enhancement module provided by the present invention; Figure 5 It is a comparison diagram of the fishery resources statistical image before and after enhancement provided by the present invention. DETAILED DESCRIPTION

[0024] The present invention provides an image enhancement processing method for fishery resource statistics, which aims to construct a color prior feature enhancement network, wherein the color prior feature enhancement network comprises a high-frequency information enhancement module and a low-frequency information enhancement module. The high-frequency information enhancement module strengthens the edges, details and complex textures in the image, so that the detail features of the fishery resource statistics image are clearer. The low-frequency information enhancement module focuses on the precise correction of the image color, ensuring that the image color information maintains consistency and accuracy during the enhancement process, thereby better retaining the overall content and color expression of the fishery resource statistics image. A frequency information interaction module is constructed to realize the interaction between the high-frequency information and the low-frequency information. The color and global content information in the low-frequency information can be effectively integrated with the texture and detail features in the high-frequency information, thereby realizing the enhancement of the fishery resource statistics image.

[0025] See also Figure 1 As shown, an image enhancement processing method for fishery resource statistics in an embodiment of the present application.

[0026] S1. Collection of fishery resource statistical images. Install the underwater camera on the underwater robot, cruise and shoot in specific waters, obtain 2000 underwater fishery resource statistical images, annotate the collected fishery resource statistical images, and the annotated images constitute the fishery resource statistical image dataset. The 2000 fishery resource statistical image dataset is divided into 1400 training sets and 600 validation sets.

[0027] S2. Construct a color prior feature enhancement network, which consists of 20 feature information enhancement modules. The feature information enhancement modules include a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module.

[0028] Furthermore, if Figure 2 As shown, input fishery resource statistics image , 640, 640 and 3 represent The height, width and channels of Decompose into high-frequency information and low-frequency information, , represents discrete wavelet transform, Represents the low-frequency information of the input fishery resources statistical image, including the content and color information of the input fishery resources statistical image. It represents the high-frequency information of the input fishery resource statistical image, including the detailed information of the global structure and texture. Then the high-frequency information and low-frequency information are input into the feature information enhancement module. The feature information enhancement module consists of a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module. After the low-frequency information enhancement module, , high frequency information After the high-frequency information enhancement module, , and Input to the frequency information interaction module to realize the information interaction between high-frequency information and low-frequency information. , represents the frequency information interaction module, Represents the low-frequency information after interaction, Represents the high-frequency information after interaction, and then and Add the elements together to get , represents the addition of elements, and Add the elements together to get , represents the addition of elements, and As the input of the next feature information enhancement module, after 20 feature information enhancement modules, we get and , and The final image is generated by inverse discrete wavelet transform , Represents inverse discrete wavelet transform.

[0029] S3. Construct a low-frequency information enhancement module, including SE attention mechanism, regularization, GELU activation function, fully connected layer and multi-scale color information extraction module, to realize the color correction of low-frequency information of fishery resource statistical images.

[0030] Furthermore, if Figure 3 As shown, for the low-frequency information enhancement module, the low-frequency information of the fishery resources statistical image is input , 320, 320 and 3 represent height, width and channel respectively. First, the SE attention mechanism adaptively assigns weights to different feature channels, enhances important features, suppresses unimportant features, and uses Dropout for regularization to prevent overfitting of model training. The 3×3 convolutional layer and GELU activation function work together to capture low-frequency information while improving the representation ability of the model, and obtain the optimized low-frequency information. , represents the SE attention mechanism, represents Dropout regularization, represents a 3×3 convolutional layer, Represents the GELU activation function, and then uses the multi-scale color information extraction module to obtain more comprehensive color information , Represents the multi-scale color information extraction module, and finally Input to the fully connected layer to obtain enhanced low-frequency information .

[0031] S4. Construct a high-frequency information enhancement module, including a 1×1 convolutional layer, a deep convolutional layer, an attention mechanism, and a multi-layer perceptron, to enhance the high-frequency information details of fishery resource statistical images.

[0032] Furthermore, if Figure 4 As shown in the figure, for the high-frequency information enhancement module, the high-frequency information of the fishery resources statistical image is first Perform layer normalization , 320, 320 and 3 represent height, width and channel respectively, The representation layer is normalized, and the processed features are obtained by passing through a 1×1 convolution layer and a 7×7 deep convolution layer, a 1×1 convolution layer and a 5×5 deep convolution layer, and a 1×1 convolution layer and a 3×3 deep convolution layer. , , , Represents query, key and value respectively, represents a 1×1 convolutional layer, represents a 7×7 depth convolutional layer, represents a 5×5 depth convolutional layer, represents a 3×3 deep convolutional layer, and then performs self-attention calculation, layer normalization and multi-layer perception operations on Q, K and V, calculates the weighted value of the feature, improves the model's attention to important features, and finally introduces residual connections to improve the stability of model training. , represents the self-attention mechanism, Representation layer normalization processing, represents a multi-layer perception operation, Represents addition of elements.

[0033] S5. Build a frequency information interaction module, including a 1×1 convolution layer, a 3×3 deep convolution layer, and a softmax activation function, to achieve mutual reinforcement of high- and low-frequency information.

[0034] Furthermore, the frequency information interaction module is used to promote the information interaction between low-frequency information and high-frequency information, realize the mutual reinforcement of high-frequency and low-frequency information, and input high-frequency information. and low frequency information , 320, 320 and 3 represent and Height, width and channel, high frequency information After a 1×1 convolutional layer, we get , represents a 1×1 convolutional layer, express Query, high frequency information After 1×1 convolution layer and 3×3 depth convolution layer, we get , represents a 1×1 convolutional layer, represents a 3×3 depth convolutional layer, express The value of low-frequency information After a 1×1 convolutional layer, we get , represents a 1×1 convolutional layer, express Key, low frequency information After 1×1 convolution layer and 3×3 depth convolution layer, we get , represents a 1×1 convolutional layer, represents a 3×3 depth convolutional layer, express The value of and Perform softmax normalization to get , Represents the softmax activation function, and finally and Multiply elements together to get high-frequency information after interaction , Represents element-wise multiplication, and Multiply elements together to get the low-frequency information after interaction , Represents element-wise multiplication.

[0035] S6. Processing of fishery resource statistical images: inputting the fishery resource statistical images to be processed into a color prior feature enhancement network to obtain enhanced fishery resource statistical images.

[0036] Furthermore, in step S6, a color prior feature enhancement network is developed using the Python language through the Pycharm application and implemented based on the Pytorch framework. The color prior feature enhancement network inputs the fishery resource statistical image to be processed with a resolution of 640×640×3, and starts training using a self-supervised learning method to effectively extract high-frequency and low-frequency information in the image, and realize the interaction between high-frequency and low-frequency information, and finally outputs an enhanced fishery resource statistical image.

[0037] Furthermore, if Figure 5 As shown, Figure 5 The middle left half is the statistical image of fishery resources to be processed. Figure 5The middle right half is the enhanced fishery resource statistical image after being processed by the color prior feature enhancement network, which significantly improves the clarity and recognizability of the fishery resource statistical image, and improves the accuracy of subsequent fish identification, classification and quantity statistics.

[0038] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. An image enhancement processing method for fishery resource statistics, characterized in that: The following steps are involved: S1. Collection of fishery resource statistical images: installing underwater cameras on underwater robots, cruising and shooting in specific waters, obtaining underwater fishery resource statistical images, manually adding labels to the obtained fishery resource statistical images and making a data set; S2, constructing a color prior feature enhancement network, which is composed of N feature information enhancement modules, and the feature information enhancement module includes a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module; S3, construct a low-frequency information enhancement module, including SE attention mechanism, regularization, GELU activation function, fully connected layer and multi-scale color information extraction module, to achieve low-frequency information color correction of fishery resource statistical images; S4. Construct a high-frequency information enhancement module, including a 1×1 convolution layer, a deep convolution layer, an attention mechanism, and a multi-layer perceptron, to enhance the high-frequency information details of fishery resource statistical images; S5. Build a frequency information interaction module, including 1×1 convolution layer, 3×3 deep convolution layer, and softmax activation function, to achieve mutual reinforcement of high and low frequency information; S6. Processing of fishery resource statistical images: inputting the fishery resource statistical images to be processed into a color prior feature enhancement network to obtain enhanced fishery resource statistical images.

2. The image enhancement processing method for fishery resource statistics according to claim 1 is characterized in that: In S2, for the color prior feature enhancement network, the fishery resources statistical image is input , H, W and C represent The height, width and channels of Decompose into high-frequency information and low-frequency information, , represents discrete wavelet transform, Represents the low-frequency information of the input fishery resources statistical image, including the content and color information of the input fishery resources statistical image. It represents the high-frequency information of the input fishery resource statistical image, including the detailed information of the global structure and texture. Then the high-frequency information and low-frequency information are input into the feature information enhancement module. The feature information enhancement module consists of a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module. After the low-frequency information enhancement module, , high frequency information After the high-frequency information enhancement module, , and Input to the frequency information interaction module to realize the information interaction between high-frequency information and low-frequency information. , represents the frequency information interaction module, Represents the low-frequency information after interaction, Represents the high-frequency information after interaction, and then and Add the elements together to get , represents the addition of elements, and Add the elements together to get , represents the addition of elements, and As the input of the next feature information enhancement module, after N feature information enhancement modules, and , and The final image is generated by inverse discrete wavelet transform , Represents inverse discrete wavelet transform.

3. The image enhancement processing method for fishery resource statistics according to claim 1 is characterized in that: In S3, for the low-frequency information enhancement module, the low-frequency information of the fishery resources statistical image is input , H, W and C represent height, width and channel respectively. First, the SE attention mechanism adaptively assigns weights to different feature channels, enhances important features, suppresses unimportant features, and uses Dropout for regularization to prevent overfitting of model training. The 3×3 convolutional layer and GELU activation function work together to capture low-frequency information while improving the representation ability of the model, and obtain the optimized low-frequency information. , represents the SE attention mechanism, represents Dropout regularization, represents a 3×3 convolutional layer, Represents the GELU activation function, and then uses the multi-scale color information extraction module to obtain more comprehensive color information , Represents the multi-scale color information extraction module, and finally Input to the fully connected layer to obtain enhanced low-frequency information .

4. The image enhancement processing method for fishery resource statistics according to claim 1 is characterized in that: In S4, for the high-frequency information enhancement module, firstly, the high-frequency information of the fishery resources statistical image is enhanced. Perform layer normalization , H, W and C represent height, width and channel respectively, The representation layer is normalized, and the processed features are obtained by passing through a 1×1 convolution layer and a 7×7 deep convolution layer, a 1×1 convolution layer and a 5×5 deep convolution layer, and a 1×1 convolution layer and a 3×3 deep convolution layer. , , , Represents query, key and value respectively, represents a 1×1 convolutional layer, represents a 7×7 depth convolutional layer, represents a 5×5 depth convolutional layer, represents a 3×3 deep convolutional layer, and then performs self-attention calculation, layer normalization and multi-layer perception operations on Q, K and V, calculates the weighted value of the feature, improves the model's attention to important features, and finally introduces residual connections to improve the stability of model training. , represents the self-attention mechanism, Representation layer normalization processing, represents a multi-layer perception operation, Represents addition of elements.

5. The image enhancement processing method for fishery resource statistics according to claim 1 is characterized in that: In S5, the frequency information interaction module is used to promote the information interaction between low-frequency information and high-frequency information, realize the mutual reinforcement of high-frequency and low-frequency information, and input high-frequency information. and low frequency information , H, W and C represent height, width and channel respectively, high frequency information After a 1×1 convolutional layer, we get , represents a 1×1 convolutional layer, express Query, high frequency information After 1×1 convolution layer and 3×3 depth convolution layer, we get , represents a 1×1 convolutional layer, represents a 3×3 depth convolutional layer, express The value of low-frequency information After a 1×1 convolutional layer, we get , represents a 1×1 convolutional layer, express Key, low frequency information After 1×1 convolution layer and 3×3 depth convolution layer, we get , represents a 1×1 convolutional layer, represents a 3×3 depth convolutional layer, express The value of and Perform softmax normalization to get , Represents the softmax activation function, and finally and Multiply elements together to get high-frequency information after interaction , Represents element-wise multiplication, and Multiply elements together to get the low-frequency information after interaction , Represents element-wise multiplication.

6. The image enhancement processing method for fishery resource statistics according to claim 3 is characterized in that: For the multi-scale color information extraction module, in order to process and extract the information of each color channel separately, the input fishery resources statistical image features are firstly Extract the R, G, and B color channels respectively. , It represents the separation of color channels, uses 3×3 convolutional layers and ReLU activation functions to extract local features in color channels, and then uses the SE attention mechanism to assign different weights to the R, G, and B color channels to enhance important features and suppress unimportant features. The color features processed by the channel attention mechanism are fused by element-wise addition, and finally the fused color features are layer-normalized to obtain multi-scale color information. , represents a 3×3 convolutional layer, represents the ReLU activation function, represents the SE attention mechanism, Representation layer normalization processing.

Images