Water body image enhancement method for dynamic monitoring of fishery environment

Abstract

The invention provides a water body image enhancement method for dynamic monitoring of a fishery environment, and relates to the field of image enhancement. A water body image enhancement model is constructed, a self-adaptive feature enhancement module is adopted, high-frequency information of a key area of a water body image is enhanced through frequency domain feature learning and information extraction, and the image quality is improved in combination with a nonlinear enhancement strategy; a detail recovery module is introduced, and enhancement factors are dynamically adjusted based on pixel local feature similarity, so that image edges are clearer; the global information understanding module is used for calculating global dependency weights among pixels and fusing remote pixel information, so that the overall information of the water body image is more balanced, and information loss caused by local enhancement is inhibited; the method is applied to the field of fishery environment monitoring, and a clear image enhancement scheme is provided for the fishery environment monitoring.

Description

Technical Field

[0001] The invention belongs to the field of image enhancement, and in particular relates to a water body image enhancement method for dynamic monitoring of fishery environment. Background Art

[0002] In fishery environmental monitoring, the dynamic changes of water bodies directly affect the growth and ecological balance of fish. Timely and accurate acquisition of water image information is crucial for fishery production management, ecological protection and water quality monitoring. Traditional fishery environmental monitoring mainly relies on manual inspections and water quality sensors, but manual monitoring has problems such as poor real-time performance, limited coverage, and susceptibility to subjective errors. Although water quality sensors can provide physical and chemical parameters, they cannot intuitively analyze fish activity, pollutant distribution and environmental changes in the water. Therefore, water monitoring methods based on underwater cameras and computer vision have gradually become a research hotspot. However, due to the complexity of the water environment, including illumination changes, light scattering and attenuation, low contrast caused by turbid water, scale changes of target objects and motion blur, traditional image processing methods are difficult to directly apply to water monitoring scenarios. Existing water image enhancement technologies still have limitations in adapting to different water environments, retaining target details, and reducing noise interference.

[0003] Current water image processing methods include image enhancement techniques based on histogram equalization, Retinex theory, and deep learning. Histogram equalization can improve image contrast, but it can easily lead to local overexposure or loss of detail. The Retinex method enhances the brightness of low-light areas by simulating the human visual model, which is suitable for some water environments, but is prone to artifacts in high-noise conditions. Deep learning methods, such as those based on convolutional neural networks (CNNs), generative adversarial networks (GANs), or visual transformers, can achieve strong feature extraction capabilities through large amounts of data training. However, due to the difficulty of labeling underwater image data in fishery environments, deep learning methods are highly dependent on high-quality training data and have high computational costs, making them difficult to deploy in real-time in fishery monitoring scenarios. In addition, existing methods are often limited to local pixel enhancement and fail to effectively integrate global information, resulting in unsatisfactory segmentation of water targets and backgrounds, affecting monitoring accuracy. Summary of the Invention

[0004] The present invention provides a water image enhancement method for dynamic monitoring of fishery environment, aiming to improve the quality of water images and enhance the details and clarity of images through deep learning models and feature extraction enhancement technology.

[0005] The present invention aims to propose a water image enhancement method for dynamic monitoring of fishery environment, which specifically includes the following steps.

[0006] S1. Collect fishery environment images captured by dynamic monitoring, perform preprocessing and image enhancement, and generate a water image dataset.

[0007] S2. An adaptive image feature enhancement module is constructed through innovative local geometric feature learning and adaptive information extraction. A dynamically adjusted frequency domain enhancement factor is introduced to enhance the high-frequency information in key areas of the image. Water images are processed through nonlinear information enhancement, and finally weighted fusion is performed.

[0008] S3. Construct a detail recovery module based on the local feature similarity of each pixel in the water image, introduce a dynamic adjustment factor, use the adjustment factor to weight the local features, and enhance the detail information.

[0009] S4. Introduce global dependency weights to construct a global information understanding module. Through global dependency weights, the information of different pixels is fused to achieve adaptive global information aggregation.

[0010] S5. Construct a water body image enhancement model, which includes an input, an image embedding, an adaptive image feature enhancement module, a detail recovery module, a global information understanding module, a fully connected layer and an output.

[0011] S6. Water image enhancement model training and testing, using the prepared water image dataset to train the model.

[0012] Preferably, in step S1, a drone is used to take pictures of water bodies in a fishery environment in a dynamic environment, and multi-view cameras on the surface and underwater are configured to collect underwater and surface images. The images cover a variety of fishery environments, including fresh water, sea water, lakes and rivers. The water body images are preprocessed and image enhancement operations are performed to create a water body image dataset, and the dataset is divided into a training set, a validation set and a test set.

[0013] Preferably, in step S2, an adaptive image feature enhancement module is constructed, specifically comprising the following steps:

[0014] Step S21: Input fishery water body image features , 、 and Represent the water body image features The height, width and number of channels of the water body image are first calculated through the local geometric features of the image to capture the texture and edge information in the image and calculate the gradient of the water body image feature X in the horizontal and vertical directions:

[0015] ;

[0016] ;

[0017] in, represents the eigenvalue at the coordinate (x, y), and Respectively represent the local gradients of water body image features in the x and y directions, indicating the edge changes of the image;

[0018] Calculate the local geometric information of image features based on local gradients :

[0019] ;

[0020] Step S22: In order to apply convolution kernels of different strengths to different image feature regions, a convolution kernel is generated by an adaptive function based on the extracted local geometric information.

[0021] Adaptive convolution kernel Through local geometric information To generate dynamically, the generation formula is:

[0022] ;

[0023] in, is the weight parameter of global learning,

[0024] Then, the generated adaptive convolution , perform convolution processing on the water image features:

[0025] ;

[0026] in, It is the water image feature after the adaptive convolution operation, n is the size of the convolution kernel, k represents the horizontal offset, and l represents the vertical offset;

[0027] Step S23: Use Fast Fourier Transform (FFT) to transform the water body image features Convert from the spatial domain to the frequency domain:

[0028] ;

[0029] It is the representation of image features in the frequency domain, which contains all the frequency information of the image.

[0030] Define dynamic adjustment frequency domain enhancement factor To control the enhancement intensity of the high frequency part of the frequency domain, the calculation formula is:

[0031] ;

[0032] in, It is the weight learned through training that controls the influence of the gradient. is the local gradient of the image feature, which indicates the intensity of the image change at that position, and b is the offset term. is the activation function,

[0033] Use enhancement factor To dynamically adjust the high-frequency information enhancement intensity in the frequency domain, the specific enhancement process is:

[0034] ;

[0035] Then use the inverse Fourier transform IFFT to restore the enhanced results in the frequency domain to the spatial domain:

[0036] ;

[0037] Step S24: In order to further enhance the contrast and details of the water body image features, the RELU nonlinear enhancement method is used to enhance the details in the image through the activation function:

[0038] ;

[0039] Finally, the image features after frequency domain enhancement are fused with the convolution image features and the image features after nonlinear enhancement to obtain the final enhanced image features. Weighted fusion is used to balance the contributions of different enhanced features. The specific formula is:

[0040] ;

[0041] and is the weight coefficient learned during the training process, 、 and Control the contributions of spatial domain convolution features, frequency domain enhancement features and nonlinear enhancement features respectively.

[0042] Preferably, in step S2, for the adaptive image feature enhancement module, the image enhancement strategy is dynamically adjusted through local geometric feature learning and adaptive information extraction, and the high-frequency information of key areas is strengthened, thereby improving the image detail retention ability; in the frequency domain, the module introduces a dynamically adjusted frequency domain enhancement factor to focus on enhancing high-frequency information, so that the edges, details and structures of the image are clearer; the nonlinear information enhancement method is used to further optimize the image contrast and enhance the recognizability of complex texture structures; through weighted fusion of features from the spatial domain, frequency domain and nonlinear enhancement, the water body image can obtain better clarity and visual quality while maintaining the natural structure, thereby improving the robustness and adaptability to environmental changes.

[0043] Preferably, in step S3, a detail recovery module is constructed, specifically comprising the following steps:

[0044] Using local neighborhood calculate The similarity between the pixel at the point and the neighboring pixels is calculated as follows:

[0045] ;

[0046] Where, is the input image feature at pixel location The value at Indicated by pixels A 3×3 local neighborhood centered on express The eigenvalue at (i, j) in the domain, To control the exponential decay coefficient of similarity, the weight of pixels far from the center is determined. Used to measure similarity, the larger the value, the more similar the pixel is to its neighboring pixels;

[0047] Based on the similarity of local features, a dynamic adjustment factor is defined to adaptively adjust the features of different regions to make the details clearer while avoiding noise caused by over-enhancement. The adjustment factor calculation formula is:

[0048] ;

[0049] Where, Represents pixels Dynamic adjustment factor, adjustment factor Value between, Represents the smoothing factor, which controls the scope of the dynamic adjustment factor;

[0050] Using Adjustment Factors Perform weighted adjustments to local features to enhance important details in water images:

[0051] ;

[0052] Where, represents the pixel value after local feature adjustment, is the weighted mean of the local neighborhood, The calculation formula is:

[0053] .

[0054] Preferably, in step S3, the detail restoration module adaptively adjusts the image enhancement strategy based on the local feature similarity of each pixel in the water body image to improve detail preservation. Specifically, by calculating the similarity between a pixel and its neighborhood, a local feature similarity metric (LFS) is constructed to measure the correlation between the pixel and its neighborhood. Based on this, a dynamic adjustment factor (DAF) is introduced to adaptively adjust the feature enhancement strength in different regions. By using the dynamic adjustment factor to perform weighted fusion of local features, key details can be enhanced while maintaining overall image consistency, thereby improving image clarity and structural fidelity, while reducing noise interference caused by over-enhancement.

[0055] Preferably, in step S4, a global information understanding module is constructed, specifically comprising the following steps:

[0056] For any two pixels and , calculate their feature similarity , the calculation method is:

[0057] ;

[0058] Where, represents the Euclidean feature distance between two pixels, which measures their similarity. Controls the degree of similarity decay,

[0059] Then calculate the global dependency weight based on the similarity score :

[0060] ;

[0061] Using global dependency weights To adjust the features of each pixel, we can get the enhanced water image features. , making it more globally aware:

[0062] .

[0063] Preferably, in step S4, for the global information understanding module, the feature similarity between any pixel pairs is first calculated, and then the global dependency weight is calculated based on the similarity distribution to weightedly fuse the pixel information; in this way, the global information can be effectively aggregated so that each pixel point not only relies on its local features, but also makes full use of the global information to improve the overall consistency of the image.

[0064] Compared with the prior art, the present invention has the following technical effects:

[0065] Through innovative local geometric feature learning and adaptive information extraction, the high-frequency information of water images can be enhanced, and the contrast and image clarity can be improved; by introducing a dynamically adjusted frequency domain enhancement factor, adaptive enhancement is performed on key areas of the image, effectively improving the structural information of the image while avoiding noise amplification caused by excessive enhancement; in addition, the present invention combines a global information understanding module to achieve adaptive fusion of different pixel information by calculating global dependency weights, ensuring the global consistency of the image and improving the overall perceptual quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of a water body image enhancement method for dynamic monitoring of fishery environment provided by the present invention.

[0067] Figure 2 This is a structural diagram of the adaptive image feature enhancement module provided by the present invention.

[0068] Figure 3 It is a structural diagram of the detail recovery module provided by the present invention.

[0069] Figure 4 This is a structural diagram of the global information understanding module provided by the present invention.

[0070] Figure 5 This is the effect diagram of the water body image before enhancement provided by the present invention.

[0071] Figure 6 This is the effect diagram of the water body image enhancement provided by the present invention. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0073] Please see the attached Figure 1-6 The present invention provides a water image enhancement method for dynamic monitoring of fishery environment.

[0074] As attached Figure 1 The present invention proposes a water image enhancement method for dynamic monitoring of fishery environment, which includes the following steps:

[0075] S1. Collect fishery environment images captured by dynamic monitoring, perform preprocessing and image enhancement, and generate a water image dataset.

[0076] Furthermore, in step S1, a drone is used to capture images of water bodies in a fishery environment in a dynamic environment. Multi-view cameras on the surface and underwater are configured to collect underwater and surface images. The images cover a variety of fishery environments, including freshwater, seawater, lakes, and rivers. The water body images are preprocessed and image enhancement operations are performed to create a water body image dataset, and the dataset is divided into a training set, a validation set, and a test set.

[0077] S2. An adaptive image feature enhancement module is constructed through innovative local geometric feature learning and adaptive information extraction. A dynamically adjusted frequency domain enhancement factor is introduced to enhance the high-frequency information in key areas of the image. Water images are processed through nonlinear information enhancement, and finally weighted fusion is performed.

[0078] Furthermore, in step S2, an adaptive image feature enhancement module is constructed, such as Figure 2 As shown, the specific steps include:

[0079] Step S21: Input fishery water body image features , 、 and Represent the water body image features The height, width and number of channels of the water body image are first calculated through the local geometric features of the image to capture the texture and edge information in the image and calculate the gradient of the water body image feature X in the horizontal and vertical directions:

[0080] ;

[0081] ;

[0082] in, represents the eigenvalue at the coordinate (x, y), and Respectively represent the local gradients of water body image features in the x and y directions, indicating the edge changes of the image;

[0083] Calculate the local geometric information of image features based on local gradients :

[0084] ;

[0085] Step S22: In order to apply convolution kernels of different strengths to different image feature regions, a convolution kernel is generated by an adaptive function based on the extracted local geometric information.

[0086] Adaptive convolution kernel Through local geometric information To generate dynamically, the generation formula is:

[0087] ;

[0088] in, is the weight parameter of global learning. In this embodiment, The initial value is set to 0.5, the value range is 0 to 1, and it is updated using the Adam optimization algorithm.

[0089] Then, the generated adaptive convolution , perform convolution processing on the water image features:

[0090] ;

[0091] in, It is the water image feature after the adaptive convolution operation, n is the size of the convolution kernel, k represents the horizontal offset, and l represents the vertical offset;

[0092] Step S23: Use Fast Fourier Transform (FFT) to transform the water body image features Convert from the spatial domain to the frequency domain:

[0093] ;

[0094] It is the representation of image features in the frequency domain, which contains all the frequency information of the image.

[0095] Define dynamic adjustment frequency domain enhancement factor To control the enhancement intensity of the high frequency part of the frequency domain, the calculation formula is:

[0096] ;

[0097] in, is the activation function, It is the weight learned through training, which controls the influence of the gradient. In this embodiment, The initial value is set to 1, the value range is -1 to 1, it is initialized using normal distribution, and updated using the Adam optimization algorithm. It is the local gradient of the image feature, indicating the intensity of the image change at that position. b is the offset term. The initial value of b is set to 0, and the value range is -0.2 to 0.2. It is initialized with a uniform distribution and updated by the Adam optimization algorithm.

[0098] Use enhancement factor To dynamically adjust the high-frequency information enhancement intensity in the frequency domain, the specific enhancement process is:

[0099] ;

[0100] Then use the inverse Fourier transform IFFT to restore the enhanced results in the frequency domain to the spatial domain:

[0101] ;

[0102] Step S24: In order to further enhance the contrast and details of the water body image features, the RELU nonlinear enhancement method is used to enhance the details in the image through the activation function:

[0103] ;

[0104] Finally, the image features after frequency domain enhancement are fused with the convolution image features and the image features after nonlinear enhancement to obtain the final enhanced image features. Weighted fusion is used to balance the contributions of different enhanced features. The specific formula is:

[0105] ;

[0106] and is the weight coefficient learned during the training process, 、 and The contributions of spatial domain convolution features, frequency domain enhancement features and nonlinear enhancement features are controlled respectively. In this embodiment, The initial value is set to 0.4, The initial value is set to 0.4, and during the update process, and The sum is less than or equal to 1.

[0107] S3. Construct a detail recovery module based on the local feature similarity of each pixel in the water image, introduce a dynamic adjustment factor, use the adjustment factor to weight the local features, and enhance the detail information.

[0108] Furthermore, in step S3, a detail recovery module is constructed, such as Figure 3 As shown, the specific steps include:

[0109] Using local neighborhood calculate The similarity between the pixel at the point and the neighboring pixels is calculated as follows:

[0110] ;

[0111] Where, is the input image feature at pixel location The value at Indicated by pixels A 3×3 local neighborhood centered on express The eigenvalue at (i, j) in the domain, To control the exponential decay coefficient of similarity, the weight of pixels far from the center is determined. In this embodiment, The value of is set to 4, Used to measure similarity, the larger the value, the more similar the pixel is to its neighboring pixels;

[0112] Based on the similarity of local features, a dynamic adjustment factor is defined to adaptively adjust the features of different regions to make the details clearer while avoiding noise caused by over-enhancement. The adjustment factor calculation formula is:

[0113] ;

[0114] Where, Represents pixels Dynamic adjustment factor, adjustment factor Value between, Represents the smoothing factor, which controls the scope of the dynamic adjustment factor. In this embodiment, The value of is set to 0.5;

[0115] Using Adjustment Factors Perform weighted adjustments to local features to enhance important details in water images:

[0116] ;

[0117] Where, represents the pixel value after local feature adjustment, is the weighted mean of the local neighborhood, The calculation formula is:

[0118] .

[0119] S4. Introduce global dependency weights to construct a global information understanding module. Through global dependency weights, the information of different pixels is fused to achieve adaptive global information aggregation.

[0120] Furthermore, in step S4, a global information understanding module is constructed, such as Figure 4 As shown, the specific steps include:

[0121] For any two pixels and , calculate their feature similarity , the calculation method is:

[0122] ;

[0123] Where, represents the Euclidean feature distance between two pixels, which measures their similarity. Control the degree of similarity attenuation. In this embodiment, The value of is set to 2.5,

[0124] Then calculate the global dependency weight based on the similarity score :

[0125] ;

[0126] Using global dependency weights To adjust the features of each pixel, we can get the enhanced water image features. , making it more globally aware:

[0127] .

[0128] S5. Construct a water body image enhancement model, which includes an input, an image embedding, an adaptive image feature enhancement module, a detail recovery module, a global information understanding module, a fully connected layer and an output.

[0129] Furthermore, for the water image enhancement model, a low-resolution water image is input , , 、 and Respectively represent In this embodiment, H=640, W=640, C=3, that is, the input image is ,Will Input into the image embedding, which contains 3×3 convolution to obtain low-resolution water image features ,Will Input into the adaptive image feature enhancement module to obtain ,Will Input to the detail recovery module and get ,Will Input into the global information understanding module to obtain enhanced water image features ,Will The input is processed by 3×3 convolution, and the processed water image features are input to the fully connected layer, and finally the output is obtained. , .

[0130] S6. Water image enhancement model training and testing, using the prepared water image dataset to train the model.

[0131] Further, in step S6, the water image enhancement model model is coded using the Pycharm application and Python language, using the Pytorch framework, and the model input is trained with a low-resolution water image with a resolution of 640×640×3, and the model is trained from scratch for 200 epochs.

[0132] Further, if Figure 5 and Figure 6 As shown, Figure 5 shows low-resolution images of water bodies, Figure 6 The high-resolution water image processed by the water image enhancement model is displayed. It can be seen that the clarity and details of the water surface ripples and fish schools have been significantly improved, and the details are clearer.

[0133] The above are only preferred embodiments of the present invention. It should be pointed out that those 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 scope of protection of the present invention.

Claims

1. A water image enhancement method for dynamic monitoring of fishery environment, characterized in that: The following steps are involved: S1. Collect fishery environment images captured by dynamic monitoring, perform preprocessing and image enhancement, and generate a water image dataset. S2. An adaptive image feature enhancement module is constructed through innovative local geometric feature learning and adaptive information extraction. Dynamic adjustment of the frequency domain enhancement factor is introduced to enhance the high-frequency information in key areas of the image. Water images are processed through nonlinear information enhancement, and finally weighted fusion is performed. S3. Construct a detail recovery module based on the local feature similarity of each pixel in the water image, introduce a dynamic adjustment factor, use the adjustment factor to weight the local features, and enhance the detail information; S4. Introduce global dependency weights to construct a global information understanding module. Through global dependency weights, information from different pixels is fused to achieve adaptive global information aggregation. S5. Construct a water body image enhancement model, wherein the water body image enhancement model includes an input, an image embedding, an adaptive image feature enhancement module, a detail recovery module, a global information understanding module, a fully connected layer, and an output; S6. Water image enhancement model training and testing, using the prepared water image dataset to train the model.

2. The method for enhancing water body images for dynamic monitoring of fishery environment according to claim 1, characterized in that: In S1, a water body image dataset is prepared. The specific method is as follows: Unmanned aerial vehicles (UAVs) are used to capture images of water bodies in fishery environments in dynamic environments. Multi-view cameras are configured on the surface and underwater to collect underwater and surface images. The images cover a variety of fishery environments, including freshwater, seawater, lakes, and rivers. The water body images are preprocessed and enhanced to create a water body image dataset, which is then divided into training, validation, and test sets.

3. The method for enhancing water body images for dynamic monitoring of fishery environment according to claim 2, characterized in that: In S2, an adaptive image feature enhancement module is constructed, and the specific method is as follows: S21. Input fishery water body image features , 、 and Represent the water body image features The height, width and number of channels of the water body image are first calculated through the local geometric features of the image to capture the texture and edge information in the image and calculate the gradient of the water body image feature X in the horizontal and vertical directions: ； ； in, represents the eigenvalue at the coordinate (x, y), and Respectively represent the local gradients of water body image features in the x and y directions, indicating the edge changes of the image; Calculate the local geometric information of image features based on local gradients : ； S22, based on the extracted local geometric information, generate the convolution kernel through the adaptive function, Adaptive convolution kernel Through local geometric information To generate dynamically, the generation formula is: ； in, is the weight parameter of global learning, Then, the adaptive convolution , perform convolution processing on the water image features: ； in, It is the water image feature after the adaptive convolution operation, n is the size of the convolution kernel, k represents the horizontal offset, and l represents the vertical offset; S23, use fast Fourier transform FFT to transform water body image features Convert from the spatial domain to the frequency domain: ； It is the representation of image features in the frequency domain, which contains all the frequency information of the image. Define dynamic adjustment frequency domain enhancement factor To control the enhancement intensity of the high frequency part of the frequency domain, the calculation formula is: ； in, are the weights learned through training, is the local gradient of the image feature, b is the offset term, is the activation function, Use enhancement factor To dynamically adjust the high-frequency information enhancement intensity in the frequency domain, the specific enhancement process is: ； Then use the inverse Fourier transform IFFT to restore the enhanced results in the frequency domain to the spatial domain: ； S24, using the RELU nonlinear enhancement method, through the activation function to enhance the details in the image: ； Finally, the image features after frequency domain enhancement are fused with the convolution image features and the image features after nonlinear enhancement to obtain the final enhanced image features. Weighted fusion is used to balance the contributions of different enhanced features. The specific formula is: ； and is the weight coefficient learned during the training process.

4. The method for enhancing water body images for dynamic monitoring of fishery environment according to claim 3, characterized in that: In S3, a detail recovery module is designed, and the specific method is as follows: Using local neighborhood calculate The similarity between the pixel at the point and the neighboring pixels is calculated as follows: ； Where, is the input image feature at pixel location The value at Indicated by pixels A 3×3 local neighborhood centered on express The eigenvalue at the point (i, j) in the domain, is the exponential decay coefficient for controlling similarity; According to the similarity of local features, a dynamic adjustment factor is defined. The calculation formula of the adjustment factor is: ； Where, Represents pixels Dynamic adjustment factor, adjustment factor Value between, Represents the smoothing factor, which controls the scope of the dynamic adjustment factor; Using Adjustment Factors Weighted adjustment of local features: ； Where, represents the pixel value after local feature adjustment, is the weighted mean of the local neighborhood, The calculation formula is: 。 5. The method for enhancing water body images for dynamic monitoring of fishery environment according to claim 4, characterized in that: In S4, a global information understanding module is designed, and the specific method is as follows: For any two pixels and , calculate their feature similarity , the calculation method is: ； Where, Controls the degree of similarity decay, Then calculate the global dependency weight based on the similarity score : ； Using global dependency weights To adjust the features of each pixel, we can get the enhanced water image features. : 。

Images