A deep learning-based image enhancement processing method for aquaculture water quality

Abstract

The application provides a fishery breeding water quality image enhancement processing method based on deep learning, high and low resolution water quality image data is collected and pretreated to generate a water quality image data set; then, a multistage recursive feature extraction and dynamic weight factor are used to construct a multistage recursive feature separation module to generate illumination and reflection components; then, high frequency information is extracted based on an adaptive Laplacian operator and fused with original features to construct a high frequency feature fusion module; a feature response adjustment module is constructed by designing a frequency domain enhancement function to perform frequency domain enhancement on the illumination component; finally, the multistage recursive feature separation, high frequency feature fusion and feature response adjustment modules are integrated to construct an image enhancement optimization model, and the model is optimized through noise loss and illumination loss; by training the deep learning model, the enhanced fishery breeding water quality image is outputted, the image quality is improved, and the water quality monitoring effect is optimized.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement, and specifically relates to a deep learning-based method for enhancing aquatic water quality images. Background Technology

[0002] Traditional aquaculture water quality monitoring mainly relies on manual sampling and laboratory analysis. This method is not only time-consuming and labor-intensive, but also difficult to achieve real-time monitoring of large areas of water. In recent years, with the development of image acquisition and sensor technologies, real-time image acquisition by cameras installed in aquaculture farms, combined with image processing technology, has enabled continuous monitoring of water quality. However, existing methods still have many limitations when facing complex underwater environments.

[0003] As the scale of aquaculture continues to expand, the complexity and management difficulty of aquaculture waters are also increasing. Although cameras and simple image processing technology can efficiently acquire a large amount of water quality image information, the image quality drops significantly when faced with adverse factors such as turbid water, uneven lighting, and suspended particulate matter. This makes it difficult to accurately assess water quality and fish health, which not only affects the accuracy of monitoring but may also lead to the failure to detect water quality problems in a timely manner, increasing the risk of fish disease outbreaks and consequently affecting the stability and economic benefits of the entire aquaculture system.

[0004] By applying the deep learning-based aquaculture water quality image enhancement processing method of this invention to aquaculture water quality images, noise can be effectively reduced, image details can be improved and enriched, and lighting conditions can be improved, thereby clearly presenting the true state of water quality. This can not only significantly improve the accuracy and efficiency of water quality monitoring and timely detect potential problems, but also provide strong support for scientific and rational aquaculture management and decision-making. It is of great significance for ensuring the healthy growth of fish, optimizing the aquaculture environment, and promoting the sustainable development of aquaculture. Summary of the Invention

[0005] This invention provides a deep learning-based method for enhancing aquaculture water quality images, aiming to improve the detail and clarity of aquaculture water quality images, so as to more accurately and efficiently detect and judge the water quality of aquaculture, and provide technical support for aquaculture water quality control.

[0006] This invention aims to propose a model for enhancing aquaculture water quality images and provide a deep learning-based method for enhancing aquaculture water quality images, comprising the following steps:

[0007] S1. Collect high-resolution and low-resolution image data of aquaculture water quality, generate a water quality image dataset, and preprocess the water quality image dataset;

[0008] S2, By analyzing the input water quality image Perform multi-level recursive feature extraction to generate image components. :

[0009] ;

[0010] ;

[0011] ;

[0012] in, Indicates the number of recursions. This represents a recursive feature generation strategy, where Indicates mixed washing of channels. This indicates max pooling. This represents depthwise separable convolution; This indicates that the input water quality image has been processed. Sub-recursive generation strategy This is represented as an input water quality image. go through The component features generated after the operation are used to construct a multi-level recursive feature separation module based on multi-level recursive feature extraction and dynamic weighting factors. Through the multi-level recursive feature separation module, the water quality image data generates illumination components and reflection components.

[0013] S3. Construct a high-frequency feature fusion module based on the adaptive Laplacian operator. Through the high-frequency feature fusion module, extract high-frequency information from the illumination component and the reflection component. The high-frequency information is fused with the original features to generate high-frequency features.

[0014] S4. Design a frequency domain enhancement strategy to construct a feature response adjustment module. Through the feature response adjustment module, the illumination component is enhanced in the frequency domain to generate the enhanced illumination component.

[0015] S5 integrates a multi-level recursive feature separation module, a high-frequency feature fusion module, and a feature response adjustment module to construct an image enhancement and optimization model for aquaculture water quality, and based on noise loss... , and light loss Define the final loss of the model ;

[0016] S6. Using the aquaculture water quality image dataset, train a deep learning-based aquaculture water quality image enhancement and optimization model, and output the enhanced aquaculture water quality image.

[0017] Preferably, in S2, constructing a multi-level recursive feature separation module specifically includes the following steps:

[0018] Step S21: By performing multi-level recursive feature extraction on the input water quality image, the image is separated into illumination and reflectance components. The specific calculation formula is as follows:

[0019] ;

[0020] ;

[0021] ;

[0022] in, This indicates the number of recursions, with each recursive operation used to extract deeper levels of feature information; Represents a recursive feature generation function, where, Indicates mixed washing of channels. This indicates max pooling. This represents depthwise separable convolution; Indicates input water quality image go through The component features obtained after each recursion.

[0023] Step S22: Based on features Calculate the dynamic weighting factor to generate the weight distribution of the illumination and reflection components. The specific calculation formula is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] in, Expressed as mean difference, Represented as variance, This is represented as a weighting factor for the illumination component. This is expressed as a weighting factor for the reflection component. and These represent the height and width of the feature map, respectively. and These represent the pixel indices in the height and width directions of the feature, respectively. Represented as the channel index of the image.

[0029] Step S23: Generate the illumination component and reflection component based on the dynamic weighting factor, and restrict the two components to a reasonable range using an activation function. The specific calculation formula is as follows:

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] in, This is represented as pixel-wise and channel-wise multiplication. Represented as Activation function, limiting component values ​​to between.

[0035] Preferably, the multi-level recursive feature separation module constructed by S2 extracts deep features of aquaculture water quality images layer by layer through multi-level recursive feature extraction. It achieves accurate separation of illumination and reflection components through dynamic weighting factors. At the same time, it uses activation functions to limit the component values ​​within a reasonable range, avoiding instability caused by excessively large or small feature value ranges. The overall design significantly improves the accuracy of separating illumination and reflection components while ensuring computational efficiency.

[0036] Preferably, in S3, constructing the high-frequency feature fusion module specifically includes the following steps:

[0037] Step S31, Illumination Components and reflection component Feature extraction is performed using three consecutive convolutional kernels, and the features are then concatenated to form the initial features. The specific calculation formula is as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] in, and These represent the height and width of the feature map, respectively. This represents the number of channels in the feature map. express Convolution operation, and These are respectively represented as the features obtained after convolution. This represents the initial feature obtained by concatenating illumination features and reflection features. This indicates a splicing operation.

[0042] Step S32, Initial Features High-frequency information is extracted based on the adaptive Laplacian operator, firstly through... The operator calculates the horizontal and vertical gradients of the image, then calculates adaptive coefficients based on the gradient magnitude and local brightness changes. The Laplacian operator is then applied to multiply these adaptive coefficients, and finally, the results are superimposed onto the initial features to obtain high-frequency information. The specific calculation formula is as follows:

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] in, and These are respectively represented as initial features exist and gradient of direction, Represented as in a specific pixel gradient magnitude, Represented as control gradient pairs The superparameter, This is represented as the variance of the local brightness of the feature map. This is represented as the mean difference in local brightness of the feature map. Represented as in a specific pixel eigenvalues, Represented as adaptive weight coefficients, for The Laplace operator result, It represents the high-frequency components obtained after applying the adaptive Laplacian operator.

[0048] Step S33: Add and fuse the adaptively enhanced high-frequency components with the initial features to generate high-frequency features. The specific calculation formula is as follows:

[0049] ;

[0050] in, This is represented as a high-frequency information fusion feature.

[0051] Preferably, the S3 high-frequency feature fusion module uses the Sobel operator to calculate the feature gradient direction, calculates the dynamic adaptive weight factor through the gradient magnitude and local brightness changes, adaptively adjusts the high-frequency enhancement effect, reduces redundant global computation, and finally adds the high-frequency components to the initial features, which not only improves the detail preservation and edge enhancement effect of the feature map, but also preserves global information. The overall design can effectively extract deep information from illumination and reflection features, and maintain the details and overall structure of the aquaculture water quality image features.

[0052] Preferably, in S4, constructing the characteristic response adjustment module specifically includes the following steps:

[0053] Step S41: Use Fast Fourier Transform to convert the low-quality illumination component to the frequency domain. The specific calculation formula is as follows:

[0054] ;

[0055] in, This is represented as a low-quality light component. Represented as Fast Fourier Transform, This is represented as the frequency domain representation of the input illumination component.

[0056] Step S42: Design and implement a frequency domain enhancement strategy to enhance high-frequency components. The specific formula is as follows:

[0057] ;

[0058] in, This is represented as a frequency domain enhancement strategy. Represented as the enhancement coefficient, This is expressed as a filter range parameter. It is expressed as the amplitude of the frequency component.

[0059] Step S43: Apply the frequency domain enhancement strategy to the frequency domain representation, and use the inverse fast Fourier transform to convert the enhanced frequency domain representation back to the spatial domain. The specific formula is as follows:

[0060] ;

[0061] ;

[0062] in, Represented as the enhanced frequency domain, This is represented as the inverse fast Fourier transform. This represents the enhanced illumination component.

[0063] Preferably, S4 uses Fast Fourier Transform and Inverse Transform to avoid the high computational complexity of direct spatial domain operations, achieving efficient frequency domain processing. At the same time, the frequency domain enhancement strategy limits the excessive amplification of low-frequency and mid-frequency components, highlighting high-frequency components while avoiding the introduction of excessive noise. The overall design can accurately enhance the details of the illumination components, avoid the limitations of traditional spatial domain processing, and improve the high-frequency information in aquaculture water quality images, making the enhancement process more accurate and efficient.

[0064] Preferably, in S5, for constructing the model loss function It integrates a multi-level recursive feature separation module, a high-frequency feature fusion module, and a feature response adjustment module, specifically including the following steps:

[0065] Step S51, Low-quality original image and high-quality truth value After multi-level recursive feature separation modules Generates low-quality illumination and reflection components. The specific calculation formula for generating high-quality illumination and high-quality reflection components is as follows:

[0066] ;

[0067] ;

[0068] in, This indicates that the operation has been performed through a multi-level recursive feature separation module.

[0069] Step S52, Low-quality illumination component and low-quality reflection components After layer-by-layer residual enhancement modules, high-quality illumination components are simultaneously... and high-quality reflection components Similarly, after passing through the layer-by-layer residual enhancement module, the specific calculation formula is as follows:

[0070] ;

[0071] ;

[0072] in, This indicates that the operation has been performed through the layer-by-layer residual enhancement module. Represented as a low-resolution feature vector. Represented as a high-resolution feature vector;

[0073] Step S53, Features and The specific calculation formulas are as follows: (This refers to the noise loss in generating low-quality images and the ground-value noise loss in generating high-quality images.)

[0074] ;

[0075] ;

[0076] in, This represents the gradient operator in the horizontal direction. This represents the gradient operator in the vertical direction. Representing second order Paradigm express Paradigm This is represented as noise loss in low-quality images. This is represented as the high-quality true-value noise loss.

[0077] Step S54 The enhanced illumination feature components are generated by the feature response adjustment fusion module, and the specific calculation formula is as follows:

[0078] ;

[0079] in, This indicates that the operation has been performed by the characteristic response adjustment module. It is represented as the illumination characteristic component.

[0080] Step S55, on the one hand, the illumination feature components Low quality light component and features Dot product is performed to obtain the final image enhancement output, while illumination feature components are also considered. and The light loss is jointly generated and calculated using the following formula:

[0081] ;

[0082] ;

[0083] in, This represents the dot product operation. Represented as the enhanced image, This is represented as light loss.

[0084] Step S56: Add the low-quality image noise loss, high-quality ground truth noise loss, and illumination loss in the model to obtain the final loss of the model training. The specific calculation formula is as follows:

[0085] ;

[0086] in, This represents the final loss calculated during model training.

[0087] In summary, compared with existing technologies, the beneficial effects of this invention, based on the adoption of this technical solution, are as follows: The deep learning-based image enhancement method for aquaculture water quality proposed in this invention, through the synergistic effect of a multi-level recursive feature separation module, a high-frequency feature fusion module, and a feature response adjustment module, achieves refined enhancement processing of low-quality water quality images. The multi-level recursive feature separation module can fully extract the illumination and reflection component features, separating detailed information in complex scenes; the high-frequency feature fusion module extracts and enhances high-frequency information through an adaptive Laplacian operator, effectively improving image clarity and edge features; the feature response adjustment module dynamically adjusts feature weights using frequency domain enhancement technology, optimizing overall visual quality and feature distribution; through the comprehensive design of these modules, the enhanced aquaculture water quality image quality becomes more stable and richer in detail. Attached Figure Description

[0088] Figure 1 This diagram illustrates the steps of a deep learning-based water quality image enhancement method for aquaculture.

[0089] Figure 2 This is a structural diagram of a multi-level recursive feature separation module.

[0090] Figure 3 This is a structural diagram of the high-frequency feature fusion module.

[0091] Figure 4 This is a structural diagram of the characteristic response adjustment module.

[0092] Figure 5 This is a structural diagram of a deep learning-based water quality image enhancement model for aquaculture.

[0093] Figure 6 Images of low-quality aquaculture water.

[0094] Figure 7 To optimize and enhance the water quality image for aquaculture. Detailed Implementation

[0095] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0096] Please see the appendix Figure 1 -Appendix Figure 6 This invention provides a deep learning-based method for enhancing aquatic water quality images.

[0097] As attached Figure 1 As shown in the flowchart, this invention proposes a deep learning-based method for enhancing aquatic water quality images. The specific implementation method includes the following steps:

[0098] S1. Collect high-resolution and low-resolution image data of aquaculture water quality, generate a water quality image dataset, and preprocess the water quality image dataset;

[0099] S2, By analyzing the input water quality image Perform multi-level recursive feature extraction to generate image components. :

[0100] ;

[0101] ;

[0102] ;

[0103] in, Indicates the number of recursions. This represents a recursive feature generation strategy, where Indicates mixed washing of channels. This indicates max pooling. This represents depthwise separable convolution; This indicates that the input water quality image has been processed. Sub-recursive generation strategy This is represented as an input water quality image. go through The component features generated after the operation are used to construct a multi-level recursive feature separation module based on multi-level recursive feature extraction and dynamic weighting factors. Through the multi-level recursive feature separation module, the water quality image data generates illumination components and reflection components.

[0104] S3. A high-frequency feature fusion module is constructed based on the adaptive Laplacian operator. Through the S3 module, high-frequency information is extracted from the illumination component and the reflection component. The high-frequency information is fused with the original features to generate high-frequency features.

[0105] S4. Design and control frequency domain enhancement strategy to construct characteristic response adjustment module. Through the S4 module, frequency domain enhancement of illumination component is performed to generate enhanced illumination component.

[0106] S5 integrates a multi-level recursive feature separation module, a high-frequency feature fusion module, and a feature response adjustment module to construct an image enhancement and optimization model for aquaculture water quality, and based on noise loss... , and light loss Define the final loss of the model ;

[0107] S6. Using the aquaculture water quality image dataset, train a deep learning-based aquaculture water quality image enhancement and optimization model, and output the enhanced aquaculture water quality image.

[0108] Further details are attached. Figure 1 S1 describes the use of a high-resolution underwater HD camera to capture high-quality ground truth images under stable shooting conditions with controlled lighting and reduced water flow interference. Secondly, a lower-resolution underwater camera is used to capture low-quality water images under poor lighting or turbid water conditions. 1200 images of each type of water quality image are captured to form a dataset for aquaculture water quality images. Each image is then uniformly processed... The size is divided into training and test sets in a 4:1 ratio.

[0109] Furthermore, as shown in the appendix Figure 1 The construction steps of the multi-level recursive feature separation module described in S2 are as follows: Figure 2 As shown, the specific implementation steps of the module include the following:

[0110] Further, step S21 involves processing the input water quality image. Multi-level recursive feature extraction is performed to generate image components. The specific calculation formula is as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] in, This indicates the number of recursions, set to 3 to balance feature extraction depth and computational complexity; Represents a recursive feature generation function, where, Indicates mixed washing of channels. This indicates max pooling, setting the pooling window size to [value]. , This represents a depthwise separable convolution, with the kernel size set to [value]. ; Indicates input water quality image go through The component features obtained after each recursion.

[0115] Further, step S22 is based on the features Calculate the dynamic weighting factor to generate the weight distribution of the illumination and reflection components. The specific calculation formula is as follows:

[0116] ;

[0117] ;

[0118] ;

[0119] ;

[0120] in, Expressed as mean difference, Represented as variance, using and Automatically adjust the weighting factors of illumination and reflection components to adapt to the feature changes of different scenes; This is represented as a weighting factor for the illumination component. This is expressed as a weighting factor for the reflection component. ; and These represent the height and width of the feature map, respectively. and These represent the pixel indices in the height and width directions of the feature, respectively. The channel index is represented as a feature.

[0121] Further, step S23 generates illumination and reflection components based on the dynamic weighting factor, and restricts the two components to a reasonable range using an activation function. The specific calculation formula is as follows:

[0122] ;

[0123] ;

[0124] ;

[0125] ;

[0126] in, This is represented as pixel-wise and channel-wise multiplication. Represented as Activation function, limiting component values ​​to This enhances the numerical stability of features and provides standardized features for subsequent feature fusion operations.

[0127] Furthermore, as shown in the appendix Figure 1 The high-frequency feature fusion module described in S3 is constructed using the following steps: Figure 3 As shown, the specific implementation steps of the module include the following:

[0128] Further, step S31 uses illumination components. and reflection component Features are extracted using three consecutive convolutional kernels and then concatenated to form the initial features. The specific calculation formula is as follows:

[0129] ;

[0130] ;

[0131] ;

[0132] in, and These represent the height and width of the feature map, respectively. This represents the number of channels in the feature map. express Convolutional operations extract multi-level features, progressively expanding the receptive field while enhancing the ability to capture local information. and These are respectively represented as the features obtained after convolution. This represents the initial feature obtained by concatenating illumination features and reflection features. This is represented as a splicing operation, which combines illumination features and reflection features to form a more complete initial feature.

[0133] Further, step S32 will use the initial features High-frequency information is extracted based on the adaptive Laplacian operator, firstly through... The operator calculates the horizontal and vertical gradients of the image, then calculates adaptive coefficients based on the gradient magnitude and local brightness changes. The Laplacian operator is then applied to multiply these adaptive coefficients, and finally, the results are superimposed onto the initial features to obtain high-frequency information. The specific calculation formula is as follows:

[0134] ;

[0135] ;

[0136] ;

[0137] ;

[0138] in, and These are respectively represented as initial features exist and gradient of direction, Represented as in a specific pixel gradient magnitude, Represented as control gradient pairs The superparameter, This is represented as the variance of the local brightness of the feature map. This is represented as the mean difference in local brightness of the feature map. Represented as in a specific pixel eigenvalues, Represented as adaptive weight coefficients, for The Laplace operator result, It represents the high-frequency components obtained after applying the adaptive Laplacian operator.

[0139] Furthermore, step S33 involves adding and fusing the adaptively enhanced high-frequency components with the initial features to generate high-frequency features. The specific calculation formula is as follows:

[0140] ;

[0141] in, Represented as high-frequency information fusion features, Preserved initial features At the same time, the high-frequency components were enhanced. The representation of details.

[0142] Furthermore, as shown in the appendix Figure 1 The construction steps of the feature response modulation module described in S4 are as follows: Figure 4 As shown, the specific implementation steps of the module include the following steps:

[0143] Further, step S41 uses Fast Fourier Transform to convert the low-quality illumination component to the frequency domain, and the specific calculation formula is as follows:

[0144] ;

[0145] in, This is represented as a low-quality light component. Represented as Fast Fourier Transform, This is represented as the frequency domain representation of the input illumination component.

[0146] Furthermore, step S42 designs and implements a frequency domain enhancement strategy to enhance high-frequency components, with the specific formula as follows:

[0147] ;

[0148] in, This is represented as a frequency domain enhancement strategy. Represented as the enhancement coefficient, This is expressed as a filter range parameter. It is expressed as the amplitude of the frequency component.

[0149] Further, step S43 applies the frequency domain enhancement strategy to the frequency domain representation and uses the inverse fast Fourier transform to convert the enhanced frequency domain representation back to the spatial domain, as shown in the following formula:

[0150] ;

[0151] ;

[0152] in, Represented as the enhanced frequency domain, This is represented as the inverse fast Fourier transform. This represents the enhanced illumination component.

[0153] Furthermore, as shown in the appendix Figure 1 The specific construction steps of the module for constructing the image enhancement and optimization model for aquaculture water quality, as described in S5, are attached. Figure 5 As shown, the specific implementation steps of the module include the following:

[0154] Furthermore, in step S51, the low-quality original image and high-quality truth value After multi-level recursive feature separation modules Generates low-quality illumination and reflection components. The specific calculation formula for generating high-quality illumination and high-quality reflection components is as follows:

[0155] ;

[0156] ;

[0157] in, This indicates that the feature separation module has been operated through multiple levels of recursion, with the recursion level set to 3.

[0158] Furthermore, in step S52, the low-quality illumination component... and low-quality reflection components After layer-by-layer residual enhancement modules, high-quality illumination components are simultaneously... and high-quality reflection components Similarly, after passing through the layer-by-layer residual enhancement module, the specific calculation formula is as follows:

[0159] ;

[0160] ;

[0161] in, This indicates that the operation has been performed through the layer-by-layer residual enhancement module. Represented as a low-resolution feature vector. Represented as a high-resolution feature vector;

[0162] Furthermore, the features in step S53 and The specific calculation formulas are as follows: (This refers to the noise loss in generating low-quality images and the ground-value noise loss in generating high-quality images.)

[0163] ;

[0164] ;

[0165] in, This represents the gradient operator in the horizontal direction. The gradient operator in the vertical direction is used in Sobel filters. Calculate the horizontal and vertical gradients. Representing second order Paradigm express Paradigm This is represented as noise loss in low-quality images. This is represented as the high-quality true-value noise loss.

[0166] Furthermore, in step S54 The enhanced illumination feature components are generated through the feature response adjustment module, and the specific calculation formula is as follows:

[0167] ;

[0168] in, This indicates that the operation has been performed by the characteristic response adjustment module. It is represented as the illumination characteristic component.

[0169] Furthermore, in step S55, one aspect is the illumination feature component. Low quality light component and features Dot product is performed to obtain the final image enhancement output, while illumination feature components are also considered. and The light loss is jointly generated and calculated using the following formula:

[0170] ;

[0171] ;

[0172] in, This represents the dot product operation. Represented as the enhanced image, This is represented as light loss.

[0173] Furthermore, in step S56, the low-quality image noise loss, high-quality ground truth noise loss, and illumination loss in the model are added together to obtain the final loss of the model training. The specific calculation formula is as follows:

[0174] ;

[0175] in, This represents the final loss calculated during model training.

[0176] Furthermore, as shown in the appendix Figure 1 The training of the deep learning-based aquaculture water quality image enhancement and optimization model described in S6 outputs enhanced aquaculture water quality images, as shown in the attached image. Figure 6 and attached Figure 7 As shown, the specific implementation steps include the following:

[0177] Furthermore, in step S6, the operating system platform used for training the model was Ubuntu, the language was Python 3.9.10, and the PyTorch deep learning framework was used for training. The hardware platform was an NVIDIA RTX3090 with 24GB of video memory. The dataset contained 2400 high-quality and low-quality images of aquaculture water quality, which were divided into training and test sets in a 4:1 ratio. During training, SGD was used as the optimizer, with an initial learning rate of 0.001. Every 20 training epochs, the learning rate decayed to 0.1, for a total of 100 training epochs. Loss function; after training, low-quality aquaculture water quality images such as Figure 6 As shown, the image is input into the aquaculture water quality image enhancement and optimization model to obtain the optimized and enhanced aquaculture water quality image, as shown below. Figure 7 As shown.

[0178] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A deep learning-based method for enhancing aquatic water quality images, characterized in that, Includes the following steps: S1. Collect high-resolution and low-resolution image data of aquaculture water quality, generate a water quality image dataset, and preprocess the water quality image dataset; S2, By analyzing the input water quality image Perform multi-level recursive feature extraction to generate image components. : ； ； ； in, Indicates the number of recursions. This represents a recursive feature generation strategy, where Indicates mixed washing of channels. This indicates max pooling. This represents depthwise separable convolution; This indicates that the input water quality image has been processed. Sub-recursive generation strategy This is represented as an input water quality image. go through The component features generated after the operation are constructed based on multi-level recursive feature extraction and dynamic weighting factors to build a multi-level recursive feature separation module. Through the multi-level recursive feature separation module, the water quality image data generates illumination components and reflection components. S3. Construct a high-frequency feature fusion module based on the adaptive Laplacian operator. Through the high-frequency feature fusion module, extract high-frequency information from the illumination component and the reflection component. The high-frequency information is fused with the original features to generate high-frequency features. S4. Design a frequency domain enhancement strategy to construct a feature response adjustment module. Through the feature response adjustment module, the illumination component is enhanced in the frequency domain to generate the enhanced illumination component. S5 integrates a multi-level recursive feature separation module, a high-frequency feature fusion module, and a feature response adjustment module to construct an image enhancement and optimization model for aquaculture water quality, and based on noise loss... , and light loss Define the final loss of the model ; S6. Using the aquaculture water quality image dataset, train a deep learning-based aquaculture water quality image enhancement and optimization model, and output the enhanced aquaculture water quality image.

2. The method for enhancing aquaculture water quality images based on deep learning according to claim 1, characterized in that, In step S2, constructing the multi-level recursive feature separation module includes the following steps: S21, By analyzing the input water quality image Perform multi-level recursive feature extraction to generate image components. ; S22, Based on characteristics Calculate the dynamic weighting factor to generate the weight distribution of the illumination and reflection components. The specific calculation formula is as follows: ； ； ； ； in, Expressed as mean difference, Represented as variance, This is represented as a weighting factor for the illumination component. This is expressed as a weighting factor for the reflection component. and These represent the height and width of the feature map, respectively. and These represent the pixel indices in the height and width directions of the feature, respectively. Represented as the channel index of the image; S23. Based on the dynamic weighting factor, generate the illumination component and the reflection component, and restrict the two components to a reasonable range through an activation function. The specific calculation formula is as follows: ； ； ； ； in, This is represented as pixel-wise and channel-wise multiplication. Represented as Activation function, limiting component values ​​to between.

3. The method for enhancing aquaculture water quality images based on deep learning according to claim 1, characterized in that, In step S3, constructing the high-frequency feature fusion module includes the following steps: S31, Illumination Components and reflection component Features are extracted using three consecutive convolutional kernels and then concatenated to form the initial features. The specific calculation formula is as follows: ； ； ; in, and These represent the height and width of the feature map, respectively. This represents the number of channels in the feature map. express Convolution operation, and These are respectively the features obtained after convolution. This represents the initial feature obtained by concatenating illumination features and reflection features. This is represented as a splicing operation; S32, Initial Features High-frequency information is extracted based on the adaptive Laplacian operator, firstly through... The operator calculates the horizontal and vertical gradients of the image, then calculates adaptive coefficients based on the gradient magnitude and local brightness changes. The Laplacian operator is then applied to multiply these adaptive coefficients, and finally, the results are superimposed onto the initial features to obtain high-frequency information. The specific calculation formula is as follows: ; ; ; ; in, and These are respectively represented as initial features exist and Gradient of direction, Represented as in pixels gradient magnitude, Represented as control gradient pairs The superparameter, This is represented as the variance of the local brightness of the feature map. This is represented as the mean difference in local brightness of the feature map. Represented as in pixels eigenvalues, Represented as adaptive weight coefficients, for The Laplace operator result, This is represented as the high-frequency component obtained after applying the adaptive Laplacian operator; S33. The adaptively enhanced high-frequency components are added to and fused with the initial features to generate high-frequency features. The specific calculation formula is as follows: ； in, This is represented as a high-frequency information fusion feature.

4. The method for enhancing aquaculture water quality images based on deep learning according to claim 1, characterized in that, In step S4, constructing the feature response adjustment module includes the following steps: S41. Use Fast Fourier Transform to convert the low-quality illumination component to the frequency domain. The specific calculation formula is as follows: ; in, This is represented as a low-quality light component. Represented as Fast Fourier Transform, This is represented as the frequency domain representation of the input illumination component; S42. Design and implement a frequency domain enhancement strategy to enhance high-frequency components. The specific formula is as follows: ; in, This is represented as a frequency domain enhancement strategy. Represented as the enhancement coefficient, This is expressed as a filter range parameter. It is expressed as the amplitude of the frequency component; S43. Apply the frequency domain enhancement strategy to the frequency domain representation, and use the inverse fast Fourier transform to convert the enhanced frequency domain representation back to the spatial domain. The specific formula is as follows: ； ； in, Represented as the enhanced frequency domain, This is represented as the inverse fast Fourier transform. This represents the enhanced illumination component.

5. The method for enhancing aquaculture water quality images based on deep learning according to claim 1, characterized in that, In S5, the model loss function is constructed. Training a water quality enhancement and optimization model for aquaculture includes the following steps: S51, Low-quality original image and high-quality truth value After multi-level recursive feature separation modules Generating low-quality light components and low-quality reflection components , Generate high-quality illumination components and high-quality reflection components The specific calculation formula is as follows: ； ； in, This indicates that the process has involved multiple levels of recursive feature separation module operations; S52, Low-quality illumination component and low-quality reflection components After layer-by-layer residual enhancement modules, high-quality illumination components are simultaneously... and high-quality reflection components Similarly, after passing through the layer-by-layer residual enhancement module, the specific calculation formula is as follows: ； ； in, This indicates that the operation has been performed using the layer-by-layer residual enhancement module. Represented as a low-resolution feature vector. Represented as a high-resolution feature vector; S53, Features and The specific calculation formulas are as follows: (This refers to the noise loss in generating low-quality images and the ground-value noise loss in generating high-quality images.) ； ； in, This represents the gradient operator in the horizontal direction. This represents the gradient operator in the vertical direction. Representing second order Paradigm express Paradigm This is represented as noise loss in low-quality images. This is expressed as high-quality true-value noise loss; S54, The enhanced illumination feature components are generated through the deep residual feature fusion module. The specific calculation formula is as follows: ; in, This indicates that the deep residual feature fusion module has been used. Represented as illumination characteristic components; S55, one aspect of illumination characteristic components Low-quality light component and features Dot product is performed to obtain the final image enhancement output, while illumination feature components are also considered. and The light loss is jointly generated and calculated using the following formula: ; ； in, This represents the dot product operation. Represented as the enhanced image, This is expressed as light loss; S56. Add the low-quality image noise loss, high-quality ground truth noise loss, and illumination loss in the model to obtain the final loss of the model training. The specific calculation formula is as follows: ； in, This represents the final loss calculated during model training.

Images