Turbid underwater image enhancement method and equipment based on self-generative adversarial network
Through the turbid underwater image enhancement method based on the self-generated adversarial network, combined with the cascade enhancement network and the self-adversarial module, the problem of insufficient color correction and detail reconstruction of the turbid underwater image is solved, significantly improving the image quality and achieving more efficient Underwater image enhancement effect.
Patent Information
- Application Number
- CN202410940984.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2024-11-29
AI Technical Summary
The existing single underwater image enhancement method has problems such as insufficient color correction and insufficient image detail reconstruction, especially in turbid underwater environments, which are difficult to effectively improve image quality.
Using the turbid underwater image enhancement method based on the self-generated adversarial network, the data set is constructed by acquiring turbid underwater images and clear reference images, and the cascade enhancement network, color correction network and self-adversarial module are designed to perform multiple enhancements and feature reconstruction. , use adaptive standardization and normalization networks to correct color deviations, and use self-adversarial modules to compare image quality to improve enhancement effect.
Two enhancement processes for turbid underwater images were realized, which significantly improved the color shift and low contrast problems of the image, improved the performance of underwater image enhancement, and the generated images were more in line with the visual habits of the human eye, enhanced the decision-making boundary supervision, significantly Improved image detail and overall quality.
Smart Images

Figure CN119048374A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of image processing, and in particular to a method and device for enhancing turbid underwater images based on a self-generated adversarial network. Background Art
[0002] Theoretical technologies such as the acquisition, transmission and processing of underwater information are crucial to the rational development and utilization of underwater resources. Nowadays, the demand for large-scale and long-term monitoring of internal water bodies is increasing. As an important carrier and presentation form of underwater information, underwater images play an irreplaceable and important role in underwater environment detection and perception. Underwater image enhancement technology is widely used in improving seabed resource exploration, fishery monitoring, pipeline maintenance and other fields, and has been transplanted to unmanned underwater operation equipment, replacing humans in underwater monitoring operations to a certain extent.
[0003] However, affected by the special physical and chemical environment underwater, most underwater images are often affected by attenuation, color distortion and noise, resulting in low contrast, blurred details and other problems. Degraded underwater images seriously affect image enhancement and restoration, reduce imaging quality and limit application effects. In addition, underwater images have various types of color distribution, and traditional image enhancement techniques are usually ineffective.
[0004] Generative adversarial network (GAN) is widely used in underwater image enhancement (UIE) tasks. It enables the generator to learn the distribution of real sample data through continuous game between the generator and the discriminator. Due to the lack of reference images, most GAN-based UIE networks generate clearer images or images with more realistic colors by learning the mapping from synthetic underwater images to reference natural images or from original underwater images to enhanced images.
[0005] Due to the particularity of turbid water, underwater imaging often encounters problems such as attenuation, color distortion, and artificial light source noise. Existing single underwater image enhancement methods usually have two weaknesses: insufficient color correction and insufficient image detail reconstruction. Due to the lack of clear underwater images for reference, the universality of underwater image enhancement models based on generative adversarial networks is limited. Summary of the invention
[0006] The purpose of the present invention is to provide a turbid underwater image enhancement method and device based on a self-generated adversarial network in order to solve the problems of insufficient color correction and insufficient image detail reconstruction that are usually present in the existing single underwater image enhancement method.
[0007] The above-mentioned purpose of the present application is achieved through the following technical solutions:
[0008] S1: Acquire turbid underwater images and clear reference images to build an image dataset;
[0009] S2: Construct a turbid underwater image enhancement model, including: cascade enhancement network, color correction network, self-adversarial module, and feature reconstruction module;
[0010] S3: Train the turbid underwater image enhancement model using image datasets;
[0011] S4: Obtain a turbid underwater image to be reconstructed; and reconstruct the turbid underwater image to be reconstructed using the trained turbid underwater image enhancement model.
[0012] Optionally, step S3 includes:
[0013] S31: inputting the turbid underwater image into the cascade enhancement network to obtain an image after the first enhancement;
[0014] S32: standardizing and normalizing the first enhanced image through a color correction network to obtain a second enhanced image;
[0015] S33 compares the quality difference between the second enhanced image and the clear reference image through the self-adversarial module.
[0016] Optionally, the cascade enhancement network includes: an encoder module, a feature transformation module and a decoder module;
[0017] The encoder module is connected to the feature transformation module; the feature transformation module is connected to the decoder module; the feature transformation module includes: a detail enhancement attention block, a content guided attention block and an image fusion submodule; the detail enhancement attention block includes: a center difference convolution, an angle difference convolution, a horizontal difference convolution and a vertical difference convolution;
[0018] The low-level features and high-level features input by the encoder module are input into the content-guided attention block through the image fusion submodule to calculate the weights.
[0019] Optionally, the color correction network includes: a standardization network, a normalization network and a convolutional layer;
[0020] The standardized network is connected to the normalized network, and the normalized network is connected to the convolutional layer;
[0021] The standardized network consists of convolutional layers, BN modules, and sigmoid activation functions;
[0022] The normalized network consists of two SE blocks;
[0023] By standardizing the network, the histograms of the color channels are matched to make the distribution of the input features consistent;
[0024] The maximum value of the standardized input feature and the minimum value of the normalized input feature are adaptively obtained through the normalization network, and each input feature is normalized according to the distribution of the input feature.
[0025] Optionally, the self-adversarial module includes: a self-discriminator and a discriminator;
[0026] The enhancement effect of the second enhanced image relative to the first enhanced image is evaluated by a self-discriminator; the training of the cascade enhancement network in the turbid underwater image enhancement model is guided according to the enhancement effect;
[0027] The discriminator compares the quality difference between the second enhanced image and the clear reference image.
[0028] Optionally, the feature reconstruction module specifically includes: a feature fusion submodule, a pooling submodule and a convolution submodule;
[0029] Through the feature fusion submodule, the input underwater image features of different scales are fused in combination with the element-wise multiplication method;
[0030] The fused features are mapped into an enhanced underwater image, that is, a reconstructed turbid underwater image, through the pooling submodule and the convolution submodule.
[0031] An electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes a turbid underwater image enhancement method based on a self-generated adversarial network.
[0032] A computer-readable storage medium stores instructions, and when the instructions are executed, a turbid underwater image enhancement method based on a self-generated adversarial network is performed.
[0033] The beneficial effects of the technical solution provided by this application are:
[0034] 1. Design two enhancement processes for the input turbid underwater image. The first enhancement extracts the feature information of the image, and the second enhancement effectively improves the color cast of the image and enhances the low contrast of the image. Design a self-adversarial discriminator to add constraints to the generation process of the enhanced image, perform more stringent decision boundary supervision, and improve the performance of underwater image enhancement.
[0035] 2. Design a parallel detail enhancement attention block containing differential convolution and basic convolution in convolutional neural networks, integrate prior information into the convolution layer to consider the low-frequency and high-frequency components of the image. It uses five convolutional layers (four differential convolutions and one original convolution) deployed in parallel for feature extraction, explicitly encoding prior information into the convolution.
[0036] 3. The self-adversarial module compares the image quality of the generated image with the corresponding high-quality image through a dual-input discriminator, constrains the enhancement by comparing the difference in image quality, and the decision boundary is located in the positive sample domain, which is opposite to the classic binary classification setting, strengthening the decision boundary constraint. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] The present application will be further described below with reference to the accompanying drawings and embodiments, in which:
[0038] Figure 1 is a flow chart of a turbid underwater image enhancement method based on a self-generated adversarial network in an embodiment of the present application;
[0039] Figure 2 It is a structural diagram of a detail enhancement attention block of a turbid underwater image enhancement method based on a self-generated adversarial network in an embodiment of the present application;
[0040] Figure 3 It is a structural diagram of a content-guided attention block of a turbid underwater image enhancement method based on a self-generated adversarial network in an embodiment of the present application;
[0041] Figure 4 It is a model principle diagram of a turbid underwater image enhancement method based on a self-generated adversarial network in an embodiment of the present application;
[0042] Figure 5 is a structural diagram of a self-adversarial module of a turbid underwater image enhancement method based on a self-generated adversarial network in an embodiment of the present application;
[0043] Figure 6 is a structural diagram of a color correction network of a turbid underwater image enhancement method based on a self-generated adversarial network in an embodiment of the present application;
[0044] Figure 7 It is a schematic diagram of the electronic device structure of the turbid underwater image enhancement method based on the self-generated adversarial network in the embodiment of the present application. DETAILED DESCRIPTION
[0045] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.
[0046] An embodiment of the present application provides a turbid underwater image enhancement method based on a self-generated adversarial network.
[0047] Please refer to Figure 1 , Figure 1 : is a flow chart of a turbid underwater image enhancement method based on a self-generated adversarial network in an embodiment of the present application, comprising:
[0048] S1: Acquire turbid underwater images and clear reference images to build an image dataset;
[0049] Specifically, the image dataset uses multiple real turbid underwater images taken in different environments and the public dataset EUVP dataset to train the proposed network. The testing process uses the UIE benchmark dataset, which contains rich underwater scenes and covers most of the turbid underwater degradation features, which is conducive to verifying the effectiveness of the enhancement method.
[0050] S2: Construct a turbid underwater image enhancement model, including: cascade enhancement network, color correction network, self-adversarial module, and feature reconstruction module;
[0051] S3: Train the turbid underwater image enhancement model using image datasets;
[0052] S4: Obtain a turbid underwater image to be reconstructed; and reconstruct the turbid underwater image to be reconstructed using the trained turbid underwater image enhancement model.
[0053] Specifically, the image dataset uses multiple turbid underwater images taken in different environments and their corresponding clear reference images. This dataset contains rich underwater scenes and covers most of the turbid underwater degradation features, which is conducive to verifying the effectiveness of the enhancement method. Figure 4 It is a model schematic diagram of the turbid water underwater image enhancement model;
[0054] Specifically, the purpose of the self-discriminator is to train a binary classifier to evaluate the enhancement effect of the underwater image after secondary enhancement, and use it to guide the training of the generator G. The discriminator is responsible for distinguishing the authenticity of the enhanced image and the reference image and giving a result.
[0055] Step S3 includes:
[0056] S31: inputting the turbid underwater image into the cascade enhancement network to obtain an image after the first enhancement;
[0057] S32: standardizing and normalizing the first enhanced image through a color correction network to obtain a second enhanced image;
[0058] S33 compares the quality difference between the second enhanced image and the clear reference image through the self-adversarial module.
[0059] Specifically, the GAN-based enhancement method achieves a balanced effect by competing between the generator and the discriminator. The generative network D realizes the extraction of underwater degraded image features and adopts a symmetric encoder-decoder structure, which consists of a downsampling module and an upsampling module.
[0060] Specifically, image standardization and image normalization are two classic color distribution correction methods. Image standardization uses the mean and standard deviation for feature scaling, and image normalization uses the maximum and minimum values of pixels to convert pixels to a similar scale. Machine learning technology is used to embody traditional feature standardization and normalization methods as data-driven networks. The adaptive standardization network extracts the histogram of matching color channels through features; the adaptive normalization network determines the minimum and maximum values of feature normalization, adjusts the brightness of the image, and enhances the contrast of the image.
[0061] The present invention first inputs the original image into the generator G to generate a high-quality image, and further uses an adaptive standardization and normalization network to match the distribution of input features, correct the distorted distribution of underwater images, and facilitate training. A self-adversarial mode is combined to add further constraints to the generation process, and the generated image is supervised by a stronger decision boundary, thereby achieving quality transfer from natural images to distorted underwater images. Finally, the self-adversarial module is used to output a binary classification, that is, whether the second enhanced image is better than the first enhanced effect, which realizes the global and detailed identification of the generated image, and is conducive to generating enhanced images that are more in line with the visual habits of the human eye.
[0062] The cascade enhancement network includes: an encoder module, a feature transformation module and a decoder module;
[0063] The encoder module is connected to the feature transformation module; the feature transformation module is connected to the decoder module; the feature transformation module includes: a detail enhancement attention block, a content guided attention block and an image fusion submodule; the detail enhancement attention block includes: a center difference convolution, an angle difference convolution, a horizontal difference convolution and a vertical difference convolution;
[0064] The low-level features and high-level features input by the encoder module are input into the content-guided attention block through the image fusion submodule to calculate the weights.
[0065] Specifically, detail-enhanced attention blocks (such as Figure 2 As shown in the figure, in addition to introducing center differential convolution (CDC) and angle differential convolution (ADC) to enhance gradient information, the horizontal differential convolution (HDC) and vertical differential convolution (VDC) based on the traditional local descriptor are re-derived to obtain intensity information. Taking horizontal differential convolution as an example, the horizontal gradient is first calculated by calculating the difference between the selected pixel pairs, and the learned kernel weights are equivalently arranged through convolution training and directly convolved with the unchanged input features.
[0066] Specifically, the content guides the attention block (e.g. Figure 3 As shown in the figure, the spatial weights of feature modulation are assigned to channels, guiding the model to focus on important areas of channels and spatial channels. Among them, channel attention recalibrates features by calculating channel vector weights, and spatial attention adaptively represents information areas by calculating spatial importance weights. The content-guided attention block obtains the spatial importance weights of each input feature channel in a coarse-to-fine manner, while fully mixing the channel attention weights and spatial attention weights to ensure information interaction. First, the channel attention weights and spatial attention weights are fused together through a simple addition calculation, and then the content of the input features is used as a guide. Each channel is rearranged in an alternating manner through a channel shuffle operation to promote the fusion of information between different channels.
[0067] Specifically, the image fusion submodule inputs the low-level features of the encoder part and the corresponding high-level features into the content-guided attention block to calculate the weights and combines them using the weighted summation method. Input features are added through jump connections to alleviate the gradient vanishing problem and simplify the learning process. Finally, the fused features are projected through a 1×1 convolutional layer to obtain the final features.
[0068] The color correction network includes: a standardization network, a normalization network and a convolutional layer;
[0069] The standardized network is connected to the normalized network, and the normalized network is connected to the convolutional layer;
[0070] The standardized network consists of convolutional layers, BN modules, and sigmoid activation functions;
[0071] The normalized network consists of two SE blocks;
[0072] By standardizing the network, the histograms of the color channels are matched to make the distribution of the input features consistent;
[0073] The maximum value of the standardized input feature and the minimum value of the normalized input feature are adaptively obtained through the normalization network, and each input feature is normalized according to the distribution of the input feature.
[0074] Specifically, the adaptive normalization network is used to match the distribution of input features so that the histogram of the matching color channel is consistent with the distribution of the given feature, which helps to correct the distorted distribution of underwater images and facilitates training; the adaptive normalization network is constructed using two compressed excitation blocks (SE blocks) and traditional feature normalization methods to adaptively obtain the maximum and minimum values tuned to the normalization of specific input features, and normalize each feature according to the distribution to enhance the image contrast, eliminate the hazy effect, and restore the brightness. Through the appropriate configuration of the above two networks, the performance of underwater image enhancement is improved. Finally, the convolution operation can ensure that the restored image has three color channels.
[0075] Specifically, the color correction technology is integrated into the color correction network (such as Figure 6 As shown in Figure 2, adaptive matching of feature distribution and normalized values is performed through training. The BN module is a batch normalization technique that can solve the gradient vanishing and gradient exploding problems by introducing learnable parameters in the convolutional neural network.
[0076] The self-adversarial module includes: a self-discriminator and a discriminator;
[0077] The enhancement effect of the second enhanced image relative to the first enhanced image is evaluated by a self-discriminator; the training of the cascade enhancement network in the turbid underwater image enhancement model is guided according to the enhancement effect;
[0078] The discriminator compares the quality difference between the second enhanced image and the clear reference image.
[0079] Specifically, the self-discriminator aims to train a binary classifier to evaluate the enhancement effect of underwater images after secondary enhancement, and use this to guide the training of the generator G. The generated image is forwarded to the generator again, and the self-adversarial discriminator discriminates the two generated images to supervise the improvement of image quality during the generation process. This strategy provides better image quality through iterative enhancement.
[0080] The feature reconstruction module specifically includes: a feature fusion submodule, a pooling submodule and a convolution submodule;
[0081] Through the feature fusion submodule, the input underwater image features of different scales are fused in combination with the element-wise multiplication method;
[0082] The fused features are mapped into an enhanced underwater image, that is, a reconstructed turbid underwater image, through the pooling submodule and the convolution submodule.
[0083] Specifically, the pooling operation of the pooling submodule is used to alleviate the problem of too high feature dimension of the convolution layer during the feature extraction process.
[0084] like Figure 1The flowchart of the implementation of the method is shown in FIG. 1 , and the specific steps of the method are as follows:
[0085] (1) The original image is input into the cascaded image enhancement network based on the detail enhancement attention block to generate the first enhanced image. It is a three-layer encoder-decoder structure, consisting of an encoder part, a feature transformation part, and a decoder part, including upsampling and downsampling operations. The downsampling operation is achieved by setting the stride value to 2 and the number of output channels to twice the number of input channels. The spatial dimension is halved and the number of channels is doubled through the convolution layer. The upsampling operation is regarded as the inverse form of the downsampling operation. The downsampling operation is implemented through the deconvolution layer. The dimensions of the first, second, and third layers are C×H×W, The value of C is set to 32. Most existing methods only transform the features of the low-resolution space, resulting in the loss of detail information. The feature extraction blocks are deployed from level 1 to level 3. For feature fusion, we fuse the features after the downsampling operation and the corresponding features before the upsampling operation through skip connections. Finally, we simply use a 3×3 convolutional layer at the end to obtain the first enhancement result.
[0086] (2) An adaptive standardization and normalization network is further used to match the distribution of input features, correct the distorted distribution of underwater images, and facilitate training. The adaptive standardization network is used to coincide with the chromaticity histogram. It uses the mean and standard deviation for feature scaling and consists of a convolutional layer, a BN module, and a sigmoid activation function. The adaptive normalization network stretches the feature distribution with a trainable percentile maximum and percentile minimum. It consists of two SE blocks with maximum and minimum pooling, respectively. SE pooling includes compression and excitation operations, which adaptively recalibrate channel feature responses by explicitly modeling the interdependence between channels. Compression aims to compress global spatial information using global average pooling; the excitation module consists of a fully connected layer, a ReLU function, a fully connected layer, and a sigmoid function in sequence, which maps the input descriptor to a set of channel weights. Finally, the normalized output feature can be calculated as follows.
[0087]
[0088] Among them, n c (x,y) represents the normalized image features; x c (i, j) represents the original input image features; Indicates the maximum weight of the adaptive normalized channel for a given input feature; Represents the minimum weight of the adaptive normalized channel for a given input feature.
[0089] (3) A self-adversarial model is combined to add further constraints to the generation process, and the generated images are supervised with stronger decision boundaries, achieving quality transfer from natural images to distorted underwater images. The self-discriminator Ds and the discriminator D are both dual-input discriminators, using the Patch-GAN structure. The two feature extraction modules in them have the same structure, including three convolutional layers with a kernel of 3 and a stride of 2. The spliced features are input into the multi-scale fusion module to obtain richer feature information. After two convolutional layers, a 7×7 feature map is obtained. The discriminator D outputs a binary classification through adaptive average pooling and two full convolutional layers (FC), that is, judging whether the second enhanced image is worse or better than the first enhanced image. Combining these two discriminators realizes the global and detail identification of the generated images, which is conducive to generating beautiful images.
[0090] (4) Use the discriminant network to output binary classification, that is, whether the second enhanced image is better than the first enhanced image, which realizes the global and detailed identification of the generated image, which is conducive to generating enhanced images that are more in line with human visual habits. The loss function is crucial to supervising the final result of deep learning-based image enhancement. The total loss of the present invention is composed of two loss functions. The entire training process includes three adversarial losses, two perceptual losses and two L1 losses. The first generation loss function L loss1 Including adversarial loss L ad1 , L1 loss L 11 and the perceptual loss L p1 .
[0091] L loss1 =L ad1 +αL 11 +βL p1
[0092]
[0093]
[0094] Where G represents the generator; D represents the discriminator; E represents the exponential with base 10. D(G(x)) represents the identification of the first enhanced image, G(x) represents the image after the first enhancement, x represents the input original image, and D(y) represents the reference high-quality image.
[0095] The second adversarial loss function uses two adversarial losses L ad2 and L ads And L1 loss L 12 and the perceptual loss L p2 to generate better quality images.
[0096] L loss2 =L ads +L ad2+αL 12 +βL p2
[0097]
[0098] Among them, D s (G(x)) represents the self-identification of the first enhancement; x i represents the i-th column of the input image data; N represents the N-th column of the data; Represents the j-th convolutional layer of VGG.
[0099] like Figure 1 As shown in the figure, the enhancement method based on the generative adversarial network (GAN) achieves a balanced effect by competing between the generator and the discriminator. The generator network D realizes the extraction of underwater degraded image features and adopts a symmetric encoder-decoder structure, which consists of a downsampling module and an upsampling module. The discriminator network G distinguishes the generated image from the reference image by learning. However, due to the lack of reference images, most GAN-based UIE networks learn the mapping from synthetic images to reference images or enhanced images, making the model only applicable to certain types of underwater environments. If the enhanced image is used as a reference, pathological constraints will be imposed. Therefore, in order to solve the shortcomings of the current solution, a self-adversarial module (such as Figure 5 As shown in Figure 2, the natural images referenced are constrained and the improvement of image quality in the generation process is supervised by comparing the quality difference between the two images.
[0100] Image standardization and image normalization are two classic color distribution correction methods. Image standardization uses the mean and standard deviation for feature scaling, which is not limited by the range and is less affected by outliers. Image normalization uses the maximum and minimum values of pixels to convert pixels to a similar scale, but it is greatly affected by outliers, and it is difficult to determine the maximum and minimum values of the effective dynamic range of the color channel. This patent uses machine learning technology to embody traditional feature standardization and normalization methods as a data-driven network, and integrates color modification distribution technology into UIE's machine learning. The adaptive standardization network matches the histogram of the color channel through feature extraction; the adaptive normalization network determines the minimum and maximum values of feature normalization, adjusts the brightness of the image, and enhances the contrast of the image.
[0101] In addition, a detail enhancement attention block is designed in the network to further improve the enhanced image quality. Detail enhancement convolution (DEConv) is designed by introducing differential convolution to integrate local descriptors into normal convolution layers. Compared with ordinary convolution, DEConv has stronger representation and generalization capabilities. Content-guided attention (CGA) is designed to assign a unique spatial importance map (SIM) to each channel, encode more useful information in the features, and further propose an image fusion method based on content-guided attention mechanism, which adaptively and effectively fuses the low-level features of the encoder part and the corresponding high-level features through the learned spatial weight modulation features.
[0102] The present application also discloses an electronic device. Figure 7 , Figure 7 The electronic device 500 may include: at least one processor 501 , at least one network interface 504 , a user interface 503 , a memory 505 , and at least one communication bus 502 .
[0103] The communication bus 502 is used to realize the connection and communication between these components.
[0104] The user interface 503 may include a display screen and a camera, and the optional user interface 503 may also include a standard wired interface and a wireless interface.
[0105] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).
[0106] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.
[0107] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.
Claims
1. A turbid underwater image enhancement method based on a self-generated adversarial network, characterized in that: The method comprises the following steps: S1: Acquire turbid underwater images and clear reference images to build an image dataset; S2: Construct a turbid underwater image enhancement model, including: cascade enhancement network, color correction network, self-adversarial module, and feature reconstruction module; S3: Train the turbid underwater image enhancement model using image datasets; S4: Obtain a turbid underwater image to be reconstructed; and reconstruct the turbid underwater image to be reconstructed using the trained turbid underwater image enhancement model.
2. The method for enhancing turbid underwater images based on a self-generated adversarial network according to claim 1, characterized in that: Step S3 includes: S31: inputting the turbid underwater image into the cascade enhancement network to obtain an image after the first enhancement; S32: standardizing and normalizing the first enhanced image through a color correction network to obtain a second enhanced image; S33 compares the quality difference between the second enhanced image and the clear reference image through the self-adversarial module.
3. The method for enhancing turbid underwater images based on a self-generated adversarial network according to claim 1, characterized in that: The cascade enhancement network includes: an encoder module, a feature transformation module and a decoder module; The encoder module is connected to the feature conversion module; the feature conversion module is connected to the decoder module; The feature transformation module includes: a detail enhancement attention block, a content guided attention block and an image fusion submodule; the detail enhancement attention block includes: a center difference convolution, an angle difference convolution, a horizontal difference convolution and a vertical difference convolution; The low-level features and high-level features input by the encoder module are input into the content-guided attention block through the image fusion submodule to calculate the weights.
4. The method for enhancing turbid underwater images based on a self-generated adversarial network according to claim 1, characterized in that: The color correction network includes: a standardization network, a normalization network and a convolutional layer; The standardized network is connected to the normalized network, and the normalized network is connected to the convolutional layer; The standardized network consists of convolutional layers, BN modules, and sigmoid activation functions; The normalized network consists of two SE blocks; By standardizing the network, the histograms of the color channels are matched to make the distribution of the input features consistent; The maximum value of the standardized input feature and the minimum value of the normalized input feature are adaptively obtained through the normalization network, and each input feature is normalized according to the distribution of the input feature.
5. The method for enhancing turbid underwater images based on a self-generated adversarial network according to claim 1, characterized in that: The self-adversarial module includes: a self-discriminator and a discriminator; The enhancement effect of the second enhanced image relative to the first enhanced image is evaluated by a self-discriminator; the training of the cascade enhancement network in the turbid underwater image enhancement model is guided according to the enhancement effect; The discriminator compares the quality difference between the second enhanced image and the clear reference image.
6. The method for enhancing turbid underwater images based on a self-generated adversarial network according to claim 1, characterized in that: The feature reconstruction module specifically includes: a feature fusion submodule, a pooling submodule and a convolution submodule; Through the feature fusion submodule, combined with the element-wise multiplication method, the input underwater image features of different scales are fused; The fused features are mapped into an enhanced underwater image, that is, a reconstructed turbid underwater image, through the pooling submodule and the convolution submodule.
7. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 6.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method steps according to any one of claims 1 to 6 are performed.
Citation Information
Patent Citations
Image processing method and device, electronic equipment and computer readable storage medium
CN112712472A
Underwater image enhancement method based on self-adversarial generative adversarial network
CN115439361A
Underwater image enhancement method and device based on improved generative adversarial network
CN118333917A
Cited By
Student classroom learning state evaluation method and system based on artificial intelligence
CN120032300A
Concrete defect detection method and device, electronic equipment and storage medium
CN120431029A
Night image enhancement method based on multidirectional information extraction and color loss optimization
CN120598834A
A dam surface underwater turbidity image enhancement method based on a generative adversarial network
CN122714273A
A dam surface underwater turbidity image enhancement method based on a generative adversarial network
CN122714273B
