A method, system, medium, and program product for underwater image enhancement based on white balance and convolutional networks.
This underwater image enhancement method, based on white balance and convolutional networks, solves the problems of detail loss and color deviation in underwater image enhancement, achieving efficient image quality improvement and is applicable to tasks such as target detection and semantic segmentation.
Patent Information
- Application Number
- CN202611093593.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-22
- Publication Date
- 2026-11-17
AI Technical Summary
Existing underwater image enhancement methods suffer from detail loss and color deviation, and have poor generalization ability.
An underwater image enhancement method based on white balance and convolutional networks is adopted. By constructing an image quality evaluation network and an enhancement network, and utilizing residual connections and feature fusion techniques, combined with white balance processing and multi-scale feature extraction, differential enhancement processing is performed.
It improves the quality of underwater images, enhances the model's adaptability to different environments, significantly improves image detail preservation and color accuracy, and supports subsequent tasks such as object detection and semantic segmentation.
Smart Images

Figure CN123118385A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing technology, specifically relating to an underwater image enhancement method, system, medium, and program product based on white balance and convolutional networks. Background Technology
[0002] As a core carrier for acquiring deep-sea information, underwater imaging technology is gradually becoming a focus of scientific research and application, with people hoping to uncover the valuable insights of the ocean depths through the analysis of these images. Underwater images, as an important information medium for humankind to explore the ocean, play an irreplaceable role in many scenarios such as marine ecological monitoring, target search, and resource exploration. However, due to the absorption and scattering effects of water on light, raw underwater images often suffer from severe color distortion, blurred details, and insufficient contrast, significantly increasing the difficulty of information analysis. For example, the turbidity of the water and uneven lighting not only restrict the operational efficiency of automated equipment such as underwater fishing robots but also make subsequent tasks extremely difficult.
[0003] With the continuous development of image processing technology and computer vision, more and more researchers are using computer vision and image processing techniques to enhance the quality of underwater images. The processed images are more conducive to human understanding and exploration of the ocean compared to the original images. Therefore, a large amount of research currently focuses on underwater image enhancement techniques. Existing methods are mainly divided into three categories: those based on physical models, those based on non-physical models, and those based on artificial neural networks. Among them, methods based on artificial neural networks currently suffer from problems such as loss of detail and color deviation, and have poor generalization ability. Therefore, there is an urgent need for an efficient and stable underwater image enhancement method that leverages the complementary semantic and detail information between multi-scale features to improve the enhancement effect of the model. Summary of the Invention
[0004] The purpose of this invention is to provide an underwater image enhancement method, system, medium, and program product based on white balance and convolutional networks, which solves the problems of detail loss and color deviation in existing underwater images.
[0005] The objective of this invention is achieved through the following technical solution:
[0006] An underwater image enhancement method based on white balance and convolutional networks, comprising the following steps:
[0007] Step 1: Use the optical camera mounted on the underwater robot to acquire optical images, construct an underwater image dataset and preprocess it, dividing the underwater images into training set, validation set and test set;
[0008] Step 2: Use the training set and validation set as input to the image quality assessment network, score the quality of the images in the training set, and classify the images in the training set into low-scoring images, medium-scoring images, and high-scoring images according to the scoring results; the image quality assessment network includes a feature extraction branch and a detail enhancement branch that perform parallel processing, and outputs a quality score after feature fusion of the outputs of the two branches;
[0009] Step 3. Construct an enhancement network based on white balance and convolutional networks, and set initial parameters; according to the scoring results in Step 2, perform differential enhancement processing on images of different levels in the training set, and input the enhanced training set images into the enhancement network for training; during training, input the validation set into the current state of the enhancement network for performance verification, and adjust the training strategy according to the verification results;
[0010] Step 4: Construct a loss function, calculate the loss value between the enhanced image output by the augmentation network and the real image, and backpropagate the loss value to update the network weights;
[0011] Step 5: Test the trained network model using the test set. Determine whether the expected results are met based on the test results. If the expected results are met, the training is complete. If the expected results are not met, return to step 3.
[0012] Furthermore, the image quality assessment network in step 2 includes dual-channel processing of detail enhancement and feature extraction, specifically:
[0013] Step 2.1: Input the image used as input to the image quality assessment network into two convolution channels for convolution operation;
[0014] Step 2.2: Input the output of the first convolution channel into multiple detail enhancement modules for detail enhancement, and obtain multiple detail enhancement output results;
[0015] Step 2.3: Use the output of the second convolutional channel as the input of the backbone feature extraction network, which contains multiple layers, and each layer outputs the corresponding layer feature map;
[0016] Step 2.4: Add the multiple detail enhancement outputs to the corresponding layer feature maps in the backbone feature extraction network element by element;
[0017] Step 2.5: Use the output of the last layer of the backbone feature extraction network as the input of the fully connected layer to output the image quality score;
[0018] Step 2.6: Based on the image quality discrimination scores, classify the images in the training set into low-scoring images, medium-scoring images, and high-scoring images.
[0019] Furthermore, the differential enhancement process in step 3 includes: performing two additional detail enhancement processes on the low-rated image, performing one additional detail enhancement process on the medium-rated image, and not performing any additional detail enhancement process on the high-rated image.
[0020] Furthermore, step 3, the enhancement network based on white balance and convolutional networks, is a hybrid architecture feature fusion network that includes downsampling paths and upsampling paths, specifically including:
[0021] Step 3.1: Perform convolution and pooling processing on the training set images after differential enhancement to obtain the initial feature map;
[0022] Step 3.2: The initial feature map is sequentially processed by a multi-level detail reconstruction module for detail enhancement and downsampling, and residual connections are performed after each level of detail reconstruction to obtain the feature map;
[0023] Step 3.3: The final first-level feature map is sequentially upsampled through a multi-level upsampling module to restore it to the preset size, and the upsampled output result is obtained;
[0024] Step 3.4: Input the upsampling output result into the fully connected layer to obtain the enhanced output image.
[0025] Furthermore, each level of the detail reconstruction module includes a detail enhancement module and a downsampling unit. The detail enhancement module is used to perform white balance processing and semantic information enhancement on the input feature map, and the downsampling module is used to reduce the feature map size.
[0026] Furthermore, the detail enhancement module includes a white balance processing layer, multiple feature mapping blocks, a pooling layer, and an ScConv module, and its workflow includes:
[0027] The input feature map is processed by a white balance processing layer and a first feature mapping block, and then pooled to output the first sub-feature map.
[0028] The input feature map is downsampled by convolution and then processed together with the first sub-feature map through a white balance processing layer and a second feature mapping block before pooling to output the second sub-feature map.
[0029] The second sub-feature map is processed by the ScConv module and the third convolutional block to output the third sub-feature map;
[0030] After adding the third sub-feature map to the first sub-feature map element by element, the sum and the first sub-feature map are processed together by the ScConv module and the fourth convolutional block to output a semantically enhanced result.
[0031] Furthermore, the loss function in step 4 is a weighted sum of a pixel-level loss function and a perceptual loss function. The pixel-level loss function is used to constrain the consistency between the enhanced image and the real image in the pixel space, and the perceptual loss function is used to constrain the consistency between the enhanced image and the real image in the semantic feature space. The pixel-level loss function is an L1 loss function, and the perceptual loss function is a distance loss function based on the deep features of the pre-trained network.
[0032] A computer system includes a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the steps of an underwater image enhancement method based on white balance and convolutional networks.
[0033] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of an underwater image enhancement method based on white balance and convolutional networks.
[0034] A computer program product includes a computer program that, when executed by a processor, implements steps of an underwater image enhancement method based on white balance and convolutional networks.
[0035] The beneficial effects of this invention are as follows:
[0036] This invention optimizes the image feature generation method. First, it uses an IQA network to classify the image, then employs residual connections to fuse information between features of different sizes. Simultaneously, it utilizes white balance (Feature Transformation) to perform feature transformation on channel features, effectively addressing issues such as color deviation and low contrast in existing underwater images. For different shooting scenarios and varying water lighting, the underwater image enhancement method of this invention can improve the model's ability to recognize different environments. In other tasks, the underwater image enhancement method of this invention can significantly improve the quality of subsequent training sets, providing reliable technical support for tasks such as object detection and semantic segmentation. Attached Figure Description
[0037] Figure 1 This is a schematic diagram of the usage process of the present invention;
[0038] Figure 2 This is a schematic diagram of the hybrid architecture feature fusion network of the present invention;
[0039] Figure 3This is a schematic diagram of the detailed reconstruction module structure of the present invention;
[0040] Figure 4 This is a schematic diagram of the upsampling module structure of the present invention;
[0041] Figure 5 This is a schematic diagram of the feature mapping block structure of the present invention;
[0042] Figure 6 This is a schematic diagram of the image quality evaluation network structure of the present invention. Detailed Implementation
[0043] The present invention will now be further described with reference to the accompanying drawings.
[0044] Combination Figure 1 This invention discloses an underwater image enhancement method based on white balance and convolutional networks, specifically including the following steps:
[0045] Step 1: The underwater robot is equipped with an optical camera to acquire optical images and construct an underwater image dataset.
[0046] Step 2: Normalize the pixel values of all images in the underwater image dataset to the [0,1] interval; crop all images to a fixed size of 256×256×3; after preprocessing, divide the entire dataset according to the following standards: Training set: used for model training; Validation set: used for validation and parameter tuning during training; Test set: used for final model performance testing. (Should the proportions of each set be added?)
[0047] Step 3: Use the training set as input to the image quality evaluation network, evaluate the image quality in the training set, assign scores, and classify the images in the training set into low-scoring images based on the scoring results. Medium score image and highly rated images .
[0048] Combination Figure 6 The image quality assessment network includes dual-channel processing of detail enhancement and feature extraction, and its specific workflow is as follows:
[0049] 3.1: The image used as input to the image quality assessment network is input into two 1×1 convolution channels for convolution operation to increase the input dimensionality.
[0050]
[0051] in, This is a 1×1 convolution operation; X is the initial input.
[0052] 3.2: The result of the first 1×1 convolution channel The input is fed into the first detail enhancement module, the second detail enhancement module, and the third detail enhancement module for detail enhancement, resulting in multiple detail enhancement outputs. , and :
[0053]
[0054] in, , and These represent the first detail enhancement module, the second detail enhancement module, and the third detail enhancement module, respectively.
[0055] 3.3: The result of the second 1×1 convolution channel As input to the backbone feature extraction network, the backbone feature extraction network adopts a ResNet structure, consisting of 3×3 convolutions with 3, 4, 6, and 3 layers respectively. The output results are denoted as follows: , , and :
[0056]
[0057] in, , , and These represent 3×3 convolutions with 3, 4, 6, and 3 layers respectively;
[0058] 3.4: The output of the last layer, consisting of 3×3 convolutions, is used as the input to the fully connected layer, and the output is the image quality judgment result. :
[0059]
[0060] in, Represented as a fully connected layer;
[0061] 3.5: Score based on image quality The underwater image dataset was divided into low-rated images. Medium score image and highly rated images Specifically, images with a quality score below the first threshold are classified as low-scoring images, images with a quality score between the first and second thresholds are classified as medium-scoring images, and images with a quality score above the second threshold are classified as high-scoring images.
[0062] This invention improves the reliability of image quality assessment by performing residual connections between multi-scale features extracted from the ResNet backbone structure and features enhanced with detail.
[0063] Step 4: Construct an underwater image enhancement network based on white balance and convolutional networks, and set initial parameters. Based on the scoring results from Step 3, perform differential enhancement processing on images of different levels in the training set, and input the enhanced training set images into the enhancement network for training. During training, input the validation set into the current state of the enhancement network for performance verification, and adjust the training strategy based on the verification results.
[0064] The differential enhancement process specifically involves: low-scoring images The medium score image serves as the input for the multi-parameter branch. High-scoring images serve as inputs to multiple parameter branches. As a fewer parameter branch; where a single multi-parameter branch module contains two additional detail enhancement processes, a single multi-parameter branch contains one additional detail enhancement process, and a single fewer parameter branch does not contain any additional detail enhancement processes;
[0065] Specifically, for low-rated images The result after 1×1 convolution and activation operations is reduced in size using global average pooling, and the low-scoring image after pooling is used as the first additional detail enhancement module. ), second additional detail enhancement module ( Enhance the quality of the input image. (This refers to a process involving input image quality enhancement.) The result after 1×1 convolution and activation operations is pooled to reduce its size, and the pooled mid-score image is used as the second detail enhancement module. The input image quality is enhanced by the first detail enhancement module and the second detail enhancement module, whose structures are similar to those of the detail enhancement module ( ). Same module:
[0066]
[0067] in , This represents the enhanced low-rated and medium-rated images. This indicates a global average pooling operation;
[0068] The enhancement network based on white balance and convolutional networks is a hybrid architecture feature fusion network that includes downsampling and upsampling paths, combined with... Figure 2 The specific workflow is as follows:
[0069] 4.1: The training set images after differential enhancement are subjected to convolution and activation processing respectively, pooling is used to reduce the size, and then 3×3 convolution is uniformly applied to obtain an initial feature map with 32 channels. :
[0070]
[0071] 4.2: Initial feature map After activation using the ReLU activation function, max pooling is used to reduce the size by half, resulting in... , initial feature map Send it to the first detailed resize block and with Perform residual connections to obtain the first feature map. First feature map The dimensions are 128×128, and the number of channels is 64.
[0072]
[0073] in, This indicates the first detail reconstruction module. This indicates a max pooling operation. This is an element-wise addition;
[0074] 4.3: The first feature map The input is fed into the second detail reconstruction module (Resize block) and the max pooling layer, and the outputs of the two are residually concatenated to obtain the second feature map. Second feature map The dimensions are 64×64, and the number of channels is 128.
[0075]
[0076] 4.4: The second feature map is input into the third detail reconstruction module (Resize block) and the max pooling layer. The outputs of the two are residually concatenated to obtain the third feature map. Third feature map The dimensions are 32×32, and the number of channels is 256.
[0077]
[0078] 4.5: The third feature map The input is fed into the first upsampling module (UPsample block) to obtain the output result. Output results The dimensions are 64×64, and the number of channels is 128;
[0079]
[0080] in, Indicates the first upsampling module;
[0081] 4.6: Output the results The input is fed into the second upsampling module (UPsample block) to obtain the output result. Output results The dimensions are 128×128, and the number of channels is 64.
[0082]
[0083] 4.7: Output the results The input is fed into the third upsampling module (UPsample block) to obtain the output result. Output results The dimensions are 256×256, and the number of channels is 32.
[0084]
[0085] 4.8: Output the results The input is fed into the first fully connected layer to obtain the trained and enhanced output. Training output The dimensions are 256×256, and the number of channels is 3.
[0086]
[0087] in, Indicates a fully connected layer;
[0088] During training, validation set images are processed using the same differential enhancement techniques as the training set and then input into the current state of the augmentation network to calculate the validation loss. The training strategy is adjusted based on the validation results, specifically including but not limited to at least one of the following methods: stopping training early to prevent overfitting when the validation loss stops decreasing or begins to increase after several consecutive rounds; dynamically adjusting the learning rate based on the trend of the validation loss; and saving the current model as the optimal model in the round with the lowest validation loss.
[0089] Furthermore, combined Figure 3 The first detail reconstruction module includes a detail enhancement module ( ) and a downsampling module, the specific workflow of which is as follows:
[0090] (1) As a detail enhancement module ( The input is used to obtain the semantically enhanced output. ;
[0091] (2) Output results with semantic information enhancement The result is used as input to the downsampling module to obtain a smaller output.
[0092] The downsampling module used in this invention can simultaneously increase the semantic information of image features, retain more detailed information of the image, and restore image color, while reducing the feature map size and memory usage.
[0093] Furthermore, combined Figure 4 The first upsampling module includes a convolutional layer, an activation function layer, and an upsampling mechanism. The specific workflow is as follows:
[0094] The third feature map The input of one convolutional layer and one activation function layer is used as the input of another convolutional layer and one activation function layer, and the output of the upsampling mechanism is used as the output of the first upsampling module. :
[0095]
[0096] in, For linear interpolation, upsampling is used; Conv is a convolutional layer.
[0097] The upsampling module of this invention can restore the image size and retain as much detail information as possible, avoiding problems such as loss of detail, blurry images, and color deviation.
[0098] Furthermore, combined Figure 3 The detail enhancement module ( It includes a first downsampling layer (DOWNsample), a first white balance processing layer (Feature Transformation), a second white balance processing layer (Feature Transformation), a first feature mapping block (Convolution block), a second feature mapping block (Convolution block), a third feature mapping block (Convolution block), a fourth feature mapping block (Convolution block), a fifth feature mapping block (Convolution block), a sixth feature mapping block (Convolution block), a max pooling mechanism, and a ScConv mechanism. The specific working process is as follows:
[0099] (1) The input is fed into a structure consisting of a first white balance processing layer, a first feature mapping block, a second feature mapping block, and a max pooling layer, and the output features are... ;
[0100]
[0101] in, This indicates the first white balance processing layer. , These represent the first feature mapping block and the second feature mapping block, respectively. The white balance processing layer specifically performs white balance processing on the feature map.
[0102] (2) After downsampling and features Simultaneously, the input is fed into a structure consisting of a second white balance processing layer, a fourth feature mapping block, a fifth feature mapping block, and a max pooling layer, and the output features are... ;
[0103]
[0104] in, This indicates the second white balance processing layer. , Indicates the fourth feature mapping block and the fifth feature mapping block;
[0105] (3) Features The input is fed into a structure consisting of ScConv and a sixth feature map block, and the output is the feature map. ;
[0106]
[0107] in, Represents the sixth feature mapping block;
[0108] (4) Features With features Add elements together, then combine the result of element-wise addition with the features. The inputs are combined into a structure consisting of ScConv and a third feature map block to obtain a semantically enhanced output. .
[0109]
[0110] in, This indicates element-wise addition.
[0111] Furthermore, combined Figure 3 The downsampling module consists of a single ordinary 3×3 convolutional layer and a single ReLU activation layer. The working process of the downsampling module is as follows:
[0112] The result The input is fed into a module consisting of a single ordinary 3×3 convolutional layer and a single ReLU activation layer, resulting in a first feature map with half the size and enhanced semantic information. :
[0113] .
[0114] Furthermore, combined Figure 5 The first feature mapping block (Convolution block) consists of two modules, each composed of a regular 3×3 convolutional layer, a dropout layer, and a ReLU activation layer. The specific workflow is as follows:
[0115] Will The input is fed into two modules consisting of a single ordinary 3×3 convolutional layer, a dropout layer, and a ReLU activation layer, resulting in a product with unchanged size and enhanced semantic information.
[0116] Step 5: Construct a new loss function, calculate the loss value between the output image of the underwater image enhancement network based on white balance and convolutional network and the real image, and backpropagate the loss value to update the weights;
[0117] The loss function is a weighted sum of a pixel-level loss function and a perceptual loss function. The pixel-level loss function is used to constrain the consistency between the enhanced image and the real image in the pixel space, and the perceptual loss function is used to constrain the consistency between the enhanced image and the real image in the semantic feature space.
[0118] loss function The calculation formula is:
[0119]
[0120] in, As a weighting factor, ;
[0121]
[0122] in, It is the original image. It is the enhanced image; The loss function is L1.
[0123]
[0124] in, Let distance be the loss function. Represents the index of a specific layer selected in the network. It is the original image. This is the enhanced image. This is the feature extraction operation of the network's layer 1, which maps the input image to the feature space of that layer. This indicates the absolute error.
[0125] This invention ensures that the output image closely approximates the target in pixel space, satisfies low-level consistency, and utilizes deep features to measure the similarity of images in semantic content. This helps to restore more natural textures and details, enabling the network to balance enhanced accuracy and visual quality.
[0126] Step 6: Test the trained network model using the test set. Determine whether the expected results are met based on the test results. If the expected results are met, the training is complete. If the expected results are not met, return to step 4.
[0127] In particular, in some preferred embodiments of the present invention, a computer device is also provided, including a memory and a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the underwater image enhancement method based on white balance and convolutional networks described in any of the above embodiments.
[0128] In some other preferred embodiments of the present invention, a computer-readable storage medium is also provided, on which a computer program / instructions are stored, which, when executed by a processor, implement the steps of the underwater image enhancement method based on white balance and convolutional networks described in any of the above embodiments.
[0129] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the underwater image enhancement method embodiments based on white balance and convolutional networks described above, which will not be repeated here.
[0130] Computer-readable storage media encompass a variety of types, including persistent and non-persistent, portable and fixed. These media store information using different technologies, and the content can be machine instructions, data structures, program modules, or other types of data. Some typical examples of computer storage media include: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), various types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory and other storage technologies, optical storage media such as CD-ROM and digital video disc (DVD), magnetic storage devices such as magnetic tape and disks, and other non-transferable media used to store information accessible to computing devices. It is important to note that the computer-readable media described herein do not include temporary storage media, such as modulated data signals and carrier waves.
[0131] Those skilled in the art will further recognize that the operation of the module can be achieved using existing technical protocols or programs, without relying on new computer programs themselves. The units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0132] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0133] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An underwater image enhancement method based on white balance and convolutional networks, characterized in that, Includes the following steps: An underwater image enhancement method based on white balance and convolutional networks, comprising the following steps: Step 1: Use the optical camera mounted on the underwater robot to acquire optical images, construct an underwater image dataset and preprocess it, dividing the underwater images into training set, validation set and test set; Step 2: Use the training set and validation set as input to the image quality assessment network, score the quality of the images in the training set, and classify the images in the training set into low-scoring images, medium-scoring images, and high-scoring images according to the scoring results; the image quality assessment network includes a feature extraction branch and a detail enhancement branch that perform parallel processing, and outputs a quality score after feature fusion of the outputs of the two branches; Step 3. Construct an enhancement network based on white balance and convolutional networks, and set initial parameters; according to the scoring results in Step 2, perform differential enhancement processing on images of different levels in the training set, and input the enhanced training set images into the enhancement network for training; during training, input the validation set into the current state of the enhancement network for performance verification, and adjust the training strategy according to the verification results; Step 4: Construct a loss function, calculate the loss value between the enhanced image output by the augmentation network and the real image, and backpropagate the loss value to update the network weights; Step 5: Test the trained network model using the test set. Determine whether the expected results are met based on the test results. If the expected results are met, the training is complete. If the expected results are not met, return to step 3.
2. The underwater image enhancement method based on white balance and convolutional networks according to claim 1, characterized in that, The image quality assessment network in step 2 includes dual-channel processing of detail enhancement and feature extraction, specifically: Step 2.1: Input the image used as input to the image quality assessment network into two convolution channels for convolution operation; Step 2.2: Input the output of the first convolution channel into multiple detail enhancement modules for detail enhancement, and obtain multiple detail enhancement output results; Step 2.3: Use the output of the second convolutional channel as the input of the backbone feature extraction network, which contains multiple layers, and each layer outputs the corresponding layer feature map; Step 2.4: Add the multiple detail enhancement outputs to the corresponding layer feature maps in the backbone feature extraction network element by element; Step 2.5: Use the output of the last layer of the backbone feature extraction network as the input of the fully connected layer to output the image quality score; Step 2.6: Based on the image quality discrimination scores, classify the images in the training set into low-scoring images, medium-scoring images, and high-scoring images.
3. The underwater image enhancement method based on white balance and convolutional networks according to claim 2, characterized in that, The differential enhancement process in step 3 includes: performing two additional detail enhancement processes on the low-rated image, performing one additional detail enhancement process on the medium-rated image, and not performing any additional detail enhancement process on the high-rated image.
4. The underwater image enhancement method based on white balance and convolutional networks according to claim 3, characterized in that, Step 3, which involves enhancing the network based on white balance and convolutional networks, is a hybrid architecture feature fusion network that includes downsampling and upsampling paths. Specifically, it includes: Step 3.1: Perform convolution and pooling processing on the training set images after differential enhancement to obtain the initial feature map; Step 3.2: The initial feature map is sequentially processed by a multi-level detail reconstruction module for detail enhancement and downsampling, and residual connections are performed after each level of detail reconstruction to obtain the feature map; Step 3.3: The final first-level feature map is sequentially upsampled through a multi-level upsampling module to restore it to the preset size, and the upsampled output result is obtained; Step 3.4: Input the upsampling output result into the fully connected layer to obtain the enhanced output image.
5. The underwater image enhancement method based on white balance and convolutional networks according to claim 4, characterized in that, Each level of the detail reconstruction module includes a detail enhancement module and a downsampling unit. The detail enhancement module is used to perform white balance processing and semantic information enhancement on the input feature map, and the downsampling module is used to reduce the size of the feature map.
6. The underwater image enhancement method based on white balance and convolutional networks according to claim 5, characterized in that, The detail enhancement module includes a white balance processing layer, multiple feature mapping blocks, a pooling layer, and an ScConv module. Its workflow includes: The input feature map is processed by a white balance processing layer and a first feature mapping block, and then pooled to output the first sub-feature map. The input feature map is downsampled by convolution and then processed together with the first sub-feature map through a white balance processing layer and a second feature mapping block before pooling to output the second sub-feature map. The second sub-feature map is processed by the ScConv module and the third convolutional block to output the third sub-feature map; After adding the third sub-feature map to the first sub-feature map element by element, the sum and the first sub-feature map are processed together by the ScConv module and the fourth convolutional block to output a semantically enhanced result.
7. The underwater image enhancement method based on white balance and convolutional networks according to claim 1, characterized in that, The loss function mentioned in step 4 is a weighted sum of a pixel-level loss function and a perceptual loss function. The pixel-level loss function is used to constrain the consistency between the enhanced image and the real image in the pixel space, and the perceptual loss function is used to constrain the consistency between the enhanced image and the real image in the semantic feature space. The pixel-level loss function is an L1 loss function, and the perceptual loss function is a distance loss function based on the deep features of the pre-trained network.
8. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 7.
10. A computer program product, comprising a computer program, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 7.
