Underwater Image Restoration Model Based on Joint Learning of Lumen Consistency and Complementary Color Correction

Abstract

This invention provides an underwater image restoration model based on joint learning of lumen consistency and complementary color correction. The invention primarily uses the geometric similarity between underwater artificial light scenes and natural light scenes to constrain brightness region consistency, guiding the model to adopt an approximate feature extraction paradigm for low-quality underwater images. It corrects the color feature drift caused by water absorption based on the color inverse relationship between complementary color images and the original low-quality image. Addressing the structural loss problem in existing underwater image enhancement methods, it constructs a phased interactive joint learning strategy based on the structural consistency characteristics of complementary color images to complete structural information. Finally, it trains and infers based on a real-world underwater dataset to restore the underwater image.

Description

Technical Field

[0001] This invention relates to the field of underwater image processing technology, and more particularly to an underwater image restoration model based on joint learning of lumen consistency and complementary color correction. Background Technology

[0002] The physical properties of the underwater environment pose a significant challenge to underwater image processing. Due to the absorption and scattering characteristics of water molecules, underwater images are often degraded. Forward scattering and backscattering are the main factors affecting image quality. Forward scattering blurs the image, while backscattering obscures image details. In addition, suspended particles in the underwater environment are also a significant factor affecting underwater image quality. These suspended particles cause scattering and absorption, leading to noise and blurring in the image. Obtaining high-quality underwater images is crucial for fields such as marine environmental monitoring, marine ecological protection, archaeology, and marine resource exploration. Therefore, researchers have developed various techniques and methods to overcome the impact of the underwater environment on image quality. For example, to address the problems of forward and backscattering, various image enhancement algorithms have been developed, such as deblurring, denoising, image restoration, and enhancement methods, to improve the clarity and visibility of underwater images. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an underwater image restoration model based on joint learning of lumen consistency and complementary color correction. This invention primarily uses the geometric similarity between underwater artificial light scenes and natural light scenes to constrain brightness region consistency, guiding the model to adopt an approximate feature extraction paradigm for low-quality underwater images. It corrects the color feature drift caused by water absorption based on the color inverse relationship between complementary color images and the original low-quality image. Addressing the structural loss problem in existing underwater image enhancement methods, it constructs a phased interactive joint learning strategy based on the structural consistency characteristics of complementary color images to complete structural information. Finally, it trains and infers based on a real-world underwater dataset to restore the underwater image.

[0004] The technical means employed in this invention are as follows:

[0005] The underwater image restoration method based on lumen consistency and complementary color correction includes the following steps:

[0006] Step S01: Obtain the underwater low-quality image dataset and the corresponding high-quality image reference dataset, and randomly divide them into training set, validation set and test set in a ratio of 8:1:1;

[0007] Step S02: Based on the geometric similarity between artificial light scenes and natural light scenes, the underwater low-quality image is divided into illuminated regions to obtain an illuminated region segmentation feature map;

[0008] Step S03: Construct a channel adaptive branch based on the RGB channel attenuation characteristics of the underwater image, and use the low-quality underwater image as input to obtain dynamic channel weights through channel feature compression extraction;

[0009] Step S04: Extract high-dimensional features from the illumination region segmentation feature map obtained in step S02 and adaptively fuse it with the channel weights obtained in step S03 to obtain a high-dimensional feature map;

[0010] Step S05: Obtain the complementary color feature map of the input underwater low-quality image;

[0011] Step S06: Based on steps S02, S03, and S04, feature extraction is performed on the complementary color feature map. Structural consistency guidance is applied to the underwater low-quality image and its complementary color feature map to obtain the global channel sharing coefficient M. r M g M b ;

[0012] Step S07: Based on the global channel sharing coefficient M obtained in step S06 r M g M b Calculate the complementary color compensation feature map and perform color drift correction on the underwater low-quality image;

[0013] Step S08: Construct a staged joint learning strategy; perform brightness region consistency guidance and structural information compensation on the underwater low-quality image;

[0014] Step S09: Construct a combined loss function and iteratively train it using the dataset obtained in step S01 as the basic unit to obtain the final high-quality restored image.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] 1. To address the issues of complex image brightness representation and local overexposure or underexposure in reconstructed images caused by the difference in light sources between underwater natural and artificial light scenes, this invention first reveals the geometric similarity of brightness features between artificial and natural light scenes, and builds a reconstruction model based on this similarity. Simultaneously considering both artificial and natural light scenes, this simplifies the complexity of the underwater reconstruction model and increases its generalization ability.

[0017] 2. To overcome the shortcomings of existing methods in representing underwater images, this invention introduces a gating mechanism and designs an efficient underwater gating module, which improves the representation learning efficiency by increasing the model's nonlinear fitting capability.

[0018] 3. To address the issues of color feature drift and structural loss in underwater images, this invention utilizes the intrinsic features of complementary color images to construct a phased interactive joint learning strategy for color drift correction and structural information completion.

[0019] Based on the above reasons, this invention has been widely adopted in the field of underwater image restoration, and is especially suitable for complex lighting scenarios such as underwater robot grasping. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the model of the present invention.

[0022] Figure 2 This image shows a comparison of the restoration results of this invention and other algorithms on images of a diver's hand. (a) represents the initial image before restoration; (b) represents the result image processed using the UDCP (Transmission Estimation in Underwater Single Images) method; (c) represents the result image processed using the WaterNet (An Underwater Image Enhancement Benchmark Dataset and Beyond) method; (d) represents the result image processed using the UWCNN (Underwaterscene prior inspired deep underwater image and video enhancement) method; (f) represents a real-world reference image; and (e) represents the restoration result image using this invention.

[0023] Figure 3This image shows a comparison of the restoration effects of this invention and other algorithms on underwater stone statue images. (a) represents the initial image before restoration; (b) represents the result image processed using the UDCP (Transmission Estimation in Underwater Single Images) method; (c) represents the result image processed using the WaterNet (An Underwater Image EnhancementBenchmark Dataset and Beyond) method; (d) represents the result image processed using the UWCNN (Underwaterscene prior inspired deep underwater image and video enhancement) method; (f) represents a real-world reference image; and (e) represents the restoration result image using this invention.

[0024] Figure 4 This image shows a comparison of the restoration results of this invention with other algorithms for shark swarm images. (a) represents the initial image before restoration; (b) represents the result image processed using the UDCP (Transmission Estimation in Underwater Single Images) method; (c) represents the result image processed using the WaterNet (An Underwater Image EnhancementBenchmark Dataset and Beyond) method; (d) represents the result image processed using the UWCNN (Underwaterscene prior inspired deep underwater image and video enhancement) method; (f) represents the real-world reference image; and (e) represents the restoration result image using this invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, 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 should fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] To verify the generalization ability of this invention to complex underwater scenes, underwater images from different scenes were selected as the test set. Simultaneously, experimental results were compared and analyzed qualitatively and quantitatively with those of the UDCP (Transmission Estimation in Underwater Single Images) algorithm, the WaterNet (An Underwater Image Enhancement Benchmark Dataset and Beyond) algorithm, and the UWCNN (Underwater scene prior inspired deep underwater image and video enhancement) algorithm. The specific steps and principles are as follows:

[0028] like Figure 1 As shown, the present invention provides an underwater image restoration model based on joint learning of luminance consistency and complementary color correction. The construction process of this model includes the following steps:

[0029] The underwater image restoration method based on lumen consistency and complementary color correction is characterized by the following steps:

[0030] Step S01: Obtain the underwater low-quality image dataset and the corresponding high-quality image reference dataset, and randomly divide them into training set, validation set and test set in a ratio of 8:1:1;

[0031] Step S02: Based on the geometric similarity between the artificial light scene and the natural light scene, the underwater low-quality image is divided into illuminated regions to obtain an illuminated region segmentation feature map; the method for calculating the illuminated region segmentation in step S02 is as follows:

[0032] H(x) c×h×1 =AAP(x);

[0033] V(x) c×1×w =AAP(x);

[0034] Region(x) c×h×w =H(x) c×h×1 @V(x) c×1×w ;

[0035] Where x represents a low-quality underwater image, h, w, and c represent the length, width, and number of channels of the input feature map, respectively; H(x) and V(x) represent the mean brightness in the horizontal and vertical directions of the input image, respectively; footnotes indicate the calculated feature dimensions; AAP represents adaptive average pooling; @ represents matrix multiplication; and Region(x) represents the illumination region segmentation result.

[0036] Step S03: Construct a channel adaptive branch based on the RGB channel attenuation characteristics of the underwater image. Using the low-quality underwater image as input, obtain dynamic channel weights through channel feature compression extraction. The channel adaptive branch construction steps in step S03 are as follows:

[0037] Step S31: Compress the underwater low-quality image along the channel dimension to obtain the channel dimension feature compression tensor;

[0038] Step S32: Construct a fully connected network to extract features from the tensors obtained in step S31;

[0039] Step S33: Use an activation function to convert the features obtained in step S32 into channel weights. The final result is expressed as follows:

[0040] CAB(x) c×1×1 =Sigmoid(FC(AAP(x)));

[0041] Where x represents the input feature, c represents the number of input feature channels, CAB(x) represents channel adaptive branching, AAP represents channel-dimensional adaptive pooling, and the Sigmoid activation function is expressed as 1 / (1+e^(-c / c)). -x ), where FC stands for fully connected layer.

[0042] Step S04: Extract high-dimensional features from the illumination region segmentation feature map obtained in step S02 and adaptively fuse it with the channel weights obtained in step S03 to obtain a high-dimensional feature map; the calculation method for high-dimensional feature extraction and channel weight fusion in step S04 is as follows:

[0043] HDFE(x)=DWC(DWC(H(x))@DWC(V(x)));

[0044] WF(x) = W × HDFE(x);

[0045] LCA(x) = PWC(WF(x));

[0046] Where x represents a low-quality underwater image, DWC represents depth-separable convolution, H(x) and V(x) represent the mean brightness in the horizontal and vertical directions of the input image, respectively, represents matrix multiplication, HDFE represents high-dimensional feature extraction, W represents the weights obtained from the channel adaptive branch, WF represents channel fusion (Weight Fusion), PWC represents pointwise convolution, and LCA represents lumen consistency attention module.

[0047] Step S05: Obtain the complementary color feature map of the input underwater low-quality image; the formula for calculating the complementary color image in step S05 is:

[0048] CCI(x) = Concat(1-Norm) 0-1 (x i i∈[1, c];

[0049] Where x represents a low-quality underwater image, Norm 0-1 This indicates that the pixel value is normalized from [0, 255] to [0, 1], c represents the total number of input channels, i represents the i-th channel of the input image, Concat represents concatenation along the channel dimension, and CCI represents complementary color image.

[0050] Step S06: Based on steps S02, S03, and S04, feature extraction is performed on the complementary color feature map. Structural consistency guidance is applied to the underwater low-quality image and its complementary color feature map to obtain the global channel sharing coefficient M. r M g M b Step S06 calculates the global channel sharing coefficient M. r M g M b The steps are as follows:

[0051] Step S61: Build a lumen consistency attention module and obtain complementary color feature maps;

[0052] Step S62: Extract high-dimensional features from the complementary color feature map using a cascaded method;

[0053] Step S63: Use an activation function to convert the high-dimensional features obtained in step S62 into weight coefficients;

[0054] The calculation formula is:

[0055] (M r M g M b ) = Sigmoid(LCA 5 (CCI(x)));

[0056] Where x represents the input feature, CCI(x) represents the complementary color feature, LCA represents the lumen consistent attention module, 5 represents the cascade depth, and Sigmoid represents 1 / (1+e) -x M r M g M b These represent the globally shared coefficient matrices for the three RGB channels.

[0057] Step S07: Based on the global channel sharing coefficient M obtained in step S06 r M g M b Calculate the complementary color compensation feature map and perform color drift correction on the underwater low-quality image;

[0058] Step S08: Construct a staged joint learning strategy; perform brightness region consistency guidance and structural information compensation on the underwater low-quality image; the steps of the staged joint learning strategy in step S08 are as follows:

[0059] Step S81: Perform initial feature extraction on the input image to obtain shallow features F1;

[0060] Step S82: Construct a 3-stage feature extraction module, each of which includes multiple sets of lumen consistency attention modules;

[0061] Step S83: Based on the output of the feature extraction module in step S82, use the color drift correction method described in step S07 to perform phased structural and color information correction, and use the correction result as the input for the next stage.

[0062] Step S09: Construct a combined loss function, and iteratively train it using the dataset obtained in step S01 as the basic unit to obtain the final high-quality restored image. According to claim 1, the underwater image restoration model based on joint learning of lumen consistency and complementary color correction is characterized in that the color drift correction formula in step S07 is:

[0063] CDC(x)=x i +M i *CCI(x i );

[0064] Where x represents a low-quality underwater image, CCI(x) represents a complementary color image, M represents a global shared coefficient matrix, i represents the i-th channel of the image and i∈{r, g, b}, and CDC represents color shift correction. The combined loss function formula in step S09 is as follows:

[0065] Loss = Loss L1 +0.2×Loss SSIM +0.1×Loss Perceptual ;

[0066] Loss L1 =|yx|;

[0067] Loss SSIM =Clamp((1-SSIM_MAP(y,x) / 2),0,1);

[0068]

[0069] Where x represents the restored image, y represents the reference image; SSIM_MAP represents the SSIM index of the restored image and the reference image, Clamp indicates that the SSIM_MAP value is greater than or equal to 0 and less than or equal to 1, and F represents the feature extraction layer of the VGG model.

[0070] Example 1

[0071] like Figure 2 As shown in the figure, this invention provides a comparison of the restoration effects of other algorithms on underwater images of divers' hands. The experimental results show that all four algorithms restored the underwater images to some extent. However, compared to the real-world reference image, the UDCP algorithm's result is less effective in correcting color shift, and the restored image still exhibits a noticeable green color cast. While the WaterNet algorithm alleviates the color shift problem, it loses background details and structural information. The UWCNN algorithm preserves background details, but the color shift in the restored image is corrected in the wrong direction, resulting in an overall yellow color cast. The algorithm of this invention better addresses the color feature shift problem compared to other algorithms, while also better preserving the details and structural information of the background region in the restored image, making it closer to the real-world reference image.

[0072] like Figure 3As shown, this invention provides a comparison of the restoration effects of other algorithms on underwater sculpture images. The experimental results show that all four algorithms restored the underwater images to some extent. However, compared to the real-world reference image, the UDCP algorithm's result is less effective in color shift correction, and the restored image still exhibits a noticeable green color cast. The WaterNet algorithm loses global brightness information, resulting in blurred background boundary information. The UWCNN algorithm, compared to WaterNet, preserves background boundary information and global brightness information to some extent, but still falls short of the reference image. The algorithm of this invention better inherits global brightness information compared to other algorithms, while also better preserving the boundary and structural information of the background region in the restored image, making it closer to the real-world reference image.

[0073] like Figure 4 As shown, this invention provides a comparison of the restoration effects of other algorithms on shark swarm images. The experimental results show that all four algorithms restored the underwater image to some extent. However, compared to the real-world reference image, the UDCP algorithm's result is less effective in color shift correction, and the restored image still exhibits a noticeable blue color cast. The WaterNet algorithm loses global brightness information, resulting in blurred background boundary information. The UWCNN algorithm, compared to WaterNet, preserves background boundary information and global brightness information to some extent, but still falls short of the reference image. The algorithm of this invention better inherits global brightness information compared to other algorithms, while also better preserving the boundary and structural information of the background region in the restored image, making it closer to the real-world reference image.

[0074] This embodiment compares the experimental results of different algorithms using two objective metrics: PSNR and SSIM. PSNR represents the ratio of the maximum possible power of a signal to the power of destructive noise that affects its representation accuracy. SSIM describes the similarity between two images in terms of brightness, contrast, and structure. As shown in Tables 1 and 2, the PSNR and SSIM scores of the UDCP, WaterNet, and UWCNN algorithms are all lower than those of the original images, but the image restored by this method shows a significant improvement over the original. This invention uses lumen consistency constraints to restore global brightness information, employs complementary color images to correct color drift, and constructs a phased joint learning strategy to complete the image structure information. Therefore, this invention significantly improves the PSNR and SSIM scores of the original image and outperforms other underwater image restoration algorithms.

[0075] Table 1. Comparison of PSNR between the model of this invention and other advanced algorithms.

[0076]

[0077] Table 2. Comparison of SSIM results between the model of this invention and other advanced algorithms.

[0078]

[0079] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0080] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0081] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0082] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0083] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0084] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An underwater image restoration method based on lumen consistency and complementary color correction, characterized in that, Comprising the following steps: Step S01: Obtain an underwater low-quality image dataset and a corresponding high-quality image reference dataset, and randomly divide the training set, the validation set and the test set according to the ratio of ​ Step S02: According to the geometric structure similarity of artificial light scene and natural light scene, the underwater low-quality image is divided into illumination area, and the illumination area division feature map is obtained; Step S03: According to the channel adaptive branch constructed according to the channel attenuation characteristics of underwater image, the underwater low-quality image is taken as input, and the dynamic channel weight is obtained by channel feature compression extraction; Step S04: High-dimensional feature extraction is performed on the illumination area division feature map obtained in step S02, and adaptive fusion is performed with the channel weight obtained in step S03, and a high-dimensional feature map is obtained; Step S05: Obtain the complementary color feature map of the input underwater low-quality image; Step S06: According to the feature extraction of the complementary color feature map in steps S02, S03 and S04, the structural consistency guidance is performed on the underwater low-quality image and its complementary color feature map, and a global channel sharing coefficient is obtained , , ; Step S07: calculating the global channel sharing coefficient according to the global channel sharing coefficient obtained in the step S06 , , calculating a complementary color compensation feature map, and performing color drift correction on the underwater low-quality image; Step S08: Construct a stage joint learning strategy; The brightness area consistency guidance and structure information compensation are performed on the underwater low-quality image; Step S09: Build a combined loss function, and take the data set obtained in step S01 as the basic unit for iterative training to obtain the final high-quality recovery image.

2. The method for underwater image restoration based on lumen consistency and complementary color correction of claim 1, wherein, The step S02 illumination area division calculation method is: ； ； ； wherein, denotes underwater low-quality images, 、 、 denote input feature map length, width and channel number, respectively; , denote input image horizontal and vertical brightness mean, respectively, and the subscript denotes the calculated feature dimension; represents adaptive average pooling, denotes matrix multiplication, denotes the light region division result.

3. The method for underwater image restoration based on lumen consistency and complementary color correction of claim 1, wherein, The step S03 channel adaptive branch construction step is: Step S31: The underwater low-quality image is compressed along the channel dimension to obtain a channel dimension feature compression tensor; Step S32: Build a fully connected network to extract features from the tensor obtained in step S31; Step S33: Use the activation function to convert the features obtained in step S32 to channel weights, and the final result is represented as: ； wherein, x represents input features, c represents input feature channel number, represents channel adaptive branching, AAP represents channel dimension adaptive pooling, activation function function expression is represented as , represents a fully connected layer.

4. The method for underwater image restoration based on lumen consistency and complementary color correction of claim 1, wherein, The high-dimensional feature extraction and channel weight fusion calculation method of step S04 is: ； ； ； wherein, represents an underwater low-quality image, represents a depthwise separable convolution, , respectively represent the horizontal and vertical brightness mean values of the input image, and represents matrix multiplication, represents high-dimensional feature extraction; represents the weight obtained by the channel adaptive branch, represents channel fusion; represents point-by-point convolution, represents a lumen consistency attention module.

5. The method for underwater image restoration based on lumen consistency and complementary color correction of claim 1, wherein, The complementary color feature map calculation formula of step S05 is: ； wherein, represents an underwater low-quality image, represents that the pixel value is normalized to , represents the total number of input channels, represents the i-th channel of the input image, represents concatenation along the channel dimension, represents a complementary color image.​​ 6. The underwater image restoration method based on joint learning of luminance consistency and complementary color correction according to claim 1, characterized in that, The step S06 calculates the global lane sharing coefficient , , The steps are: Step S61: Build a lumen consistency attention module to obtain a complementary color feature map; Step S62: Extract high-dimensional features of the complementary color feature map in a cascading manner; Step S63: Use the activation function to convert the high-dimensional features obtained in step S62 to weight coefficients; The calculation formula is: ； wherein, denotes an input feature, denotes a complementary color feature, denotes a lumen consistent attention module, 5 denotes a concatenated depth, , , , denote respectively RGB a three-channel globally shared coefficient matrix.

7. The underwater image restoration method based on joint learning of luminance consistency and complementary color correction according to claim 1, characterized in that, The color drift correction formula of step S07 is: ； wherein, represents an underwater low-quality image, represents a complementary color image, represents a global sharing coefficient matrix, represents an image channel and , represents a color drift correction.

8. The underwater image restoration method based on joint learning of luminance consistency and complementary color correction according to claim 1, characterized in that, The stage joint learning strategy step of step S08 is: Step S81: primary feature extraction is performed on the input image to obtain shallow features ; Step S82: Build a 3-stage feature extraction module, each feature extraction module includes multiple groups of lumen consistency attention modules; Step S83: According to the output result of the feature extraction module in step S82, use the color drift correction method in step S07 to correct the stage structure information and color information, and the correction result is taken as the input of the next stage.

9. The method for underwater image restoration based on lumen consistency and complementary color correction of claim 1, wherein, The combined loss function formula of step S09 is: ； ； Clamp((1 - SSIM_MAP(y, x) / 2), 0, 1); ； wherein, x denotes a restored image, denotes a reference image; denotes a difference between the restored image and the reference image, indicator, denotes a value greater than or equal to 0 and less than or equal to 1, denotes VGG a model feature extraction layer.

Images