Underwater image enhancement method of convolutional network based on underwater imaging physical model constraint

Abstract

The invention discloses an underwater image enhancement method based on a convolutional network constrained by an underwater imaging physical model, and aims to solve the problems of serious color cast and insufficient contrast, saturation and color richness of collected underwater images when an underwater robot explores an underwater world. The invention provides an underwater image enhancement method based on a convolutional network constrained by an underwater imaging physical model, and the method comprises the steps: carrying out the estimation of background light and a depth image through employing a convolutional neural network on the basis of an underwater imaging model, calculating a texture image of an image on the basis of a blurred image, and carrying out the calculation of the texture image of the image according to the texture images of different degrees of the image, and the image is enhanced from the angle of multi-scale optimization. Compared with the existing UIE algorithm, the method has the advantages that the problems of color deviation and contrast distortion are effectively eliminated while complex image details are kept, and the full-reference estimation and non-reference measurement indexes are obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine vision, and in particular relates to an underwater image enhancement method based on a convolutional network constrained by an underwater imaging physical model. Background Art

[0002] Currently, vigorously developing the marine economy and accelerating the development of the marine industry are major development priorities for countries around the world. Autonomous underwater robots, as essential tools for ocean exploration and underwater operations, play a crucial role in the development of the marine industry. High-performance underwater robots are crucial for efficient underwater operations. Visual sensors, as a crucial component of underwater robots' perception of the external environment, are crucial for improving their visual perception capabilities and addressing their inefficiency and increasing their autonomy. Underwater vision is often subject to low image quality due to environmental factors. This is primarily due to the scattering and attenuation of light by the water medium, resulting in a blue-green tint. Furthermore, due to water quality issues, captured images often exhibit low contrast and blurred details. Improving the visual quality of underwater images is crucial for the success of subsequent advanced vision tasks.

[0003] In the past, underwater image enhancement (UIE) has attracted widespread attention. Researchers have proposed a variety of underwater image enhancement methods from different perspectives. Most of these underwater enhancement methods can be divided into traditional methods and deep learning methods.

[0004] Traditional image processing methods are mainly divided into image reconstruction and image enhancement. Image reconstruction methods estimate model parameters using a prior and then invert the underwater imaging model to restore a clear image. These priors include underwater dark channel priors, minimum information loss, blur priors, attenuation curve priors, and general dark channel priors. For example, Berman et al. introduced a haze-lines model to handle wavelength-dependent attenuation and reconstruct underwater images. Furthermore, Akkaynak et al. revised the underwater imaging model based on underwater correlations. Overall, these physics-based methods are effective in certain situations, but they rely heavily on handcrafted priors. Therefore, when these priors fail in some underwater scenes, the restored image is often unsatisfactory. Image enhancement methods directly adjust the image's pixel values to achieve visually pleasing results. These methods can effectively improve the color and contrast of underwater images, but they ignore the underwater imaging mechanism. As a result, the enhanced images are prone to over-enhancement and oversaturation.

[0005] As a non-linear function approximator, neural networks have been widely used in the field of computer vision. However, they rely heavily on large datasets, and in supervised tasks, researchers often need to manually label data, which is very cumbersome. In practical engineering applications, the complex environments faced vary widely, and there is a severe lack of ground truth data for underwater images, which has become the main obstacle restricting the development of neural network models. Although researchers have synthesized a large number of underwater datasets as training samples, there are still obvious gaps between these datasets and actual problem images, and it is difficult to obtain good results as training samples. In deep learning methods, the role of underwater imaging models in underwater image enhancement is often ignored. Due to the lack of constraints from physical models, the solution space for underwater image enhancement problems is too large. Therefore, it is necessary to introduce an underwater physical imaging model into neural network training to develop a new model for underwater image enhancement. Summary of the Invention

[0006] Aiming at the problems of serious color deviation, insufficient contrast, saturation, and color richness in underwater images in the aquaculture environment, the present invention proposes an underwater image enhancement method based on a convolutional network constrained by an underwater imaging physical model. This method estimates the background light and depth map using a convolutional neural network based on the underwater imaging model, and calculates the texture map of the image based on the image blur map. According to the texture maps of different degrees of the image, the image is enhanced from the perspective of multi-scale optimization.

[0007] The technical solution to achieve the object of the present invention is: an underwater image enhancement method based on a convolutional network constrained by an underwater imaging physical model, including the following steps:

[0008] Step 1: Construct a background light estimation module to obtain the background light estimation of the input image;

[0009] Step 2: Construct a transmission map estimation module to obtain the transmission map estimation;

[0010] Step 3: Restore the underwater image using the underwater physical model;

[0011] Step 4: Solve the detailed texture map of the underwater image;

[0012] Step 5: Set the texture map weight and output the enhanced image.

[0013] An electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0014] A computer-readable storage medium, on which a computer program is stored, and the program implements the steps of the above method when executed by a processor.

[0015] Compared with the prior art, the significant advantages of the present invention are as follows: (1) The present invention can significantly improve the overall clarity of underwater images; (2) It can effectively enhance the contrast and color details of images; (3) The algorithm of the present invention shows more excellent performance in terms of average gradient (AG), underwater image quality metrics (UIQM, UCIQE), and structural similarity index measurement (SSIM and PSNR) metrics; (4) The algorithm proposed by the present invention can be combined with actual engineering cases and can be accurately and effectively used to optimize and enhance the problem of poor underwater image quality; (5) Compared with the existing UIE algorithm, this method effectively eliminates the problems of color deviation and contrast distortion while maintaining the details of complex images, and there is a significant improvement in both full-reference estimation and non-reference metric indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of the algorithm.

[0017] Figure 2 It is a schematic diagram of the background scattering estimation module.

[0018] Figure 3 It is a schematic diagram of the transmission map estimation module.

[0019] Figure 4 They are 9 representative underwater images and the images generated by the proposed method and the method of the present invention, where (a) is the original underwater image, (b) is DCP, (c) is GBdehazingRCorrection, (d) is IBLA, (e) is LowComplexityDCP, (f) is MIP, (g) is RoWS, (h) is UDCP, (i) is ULAP, and (j) is Ours. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] The present invention proposes an underwater image enhancement method based on a convolutional network constrained by an underwater imaging physical model, which estimates the background light and transmission map in the underwater imaging model by a convolutional neural network, reconstructs the input image, and the reconstructed image optimizes local details at multiple scales, thereby generating an enhanced underwater image; combined with Figure 1 , the method specifically includes the following steps:

[0022] Step 1, use the background light estimation module to calculate the estimated background light B of the input underwater image; specifically:

[0023] As Figure 2As shown, the background light estimation module mainly consists of 2 sets of convolutional kernels with a size of 5×5, 2 global average pooling layers, 2 sets of convolutional kernels with a size of 3×3, and 2 sets of convolutional kernels with a size of 1×1. The number of each set of convolutional kernels is 3. The activation function of the convolutional operation selects the Mish activation function,

[0024] f(x) = x * tanh(ln(1 + e x )) (1)

[0025] where x represents the input variable, which is the output of a certain layer in the neural network; e x is the natural exponential function, representing the exponential with the natural constant e as the base; ln(1 + e x ) is the Softplus function, which is used to smoothly map the input to the positive range; tanh(·) represents the hyperbolic tangent function, and its value range is (–1, 1). This function has no upper bound, which can prevent the network from saturating and prevent the vanishing gradient. It has a lower bound to improve the regularization effect of the network. This function is smooth and continuous at the 0 value point, which can reduce some unpredictable problems compared with ReLU. Secondly, it can make the network easier to optimize and improve the generalization performance. It can also keep some small negative inputs as negative outputs, improving the interpretability of the network and the gradient flow.

[0026] Step 2: Use the transmission map estimation module to calculate the image transmission map estimation A of the input underwater image, specifically

[0027] As Figure 3 shown, the transmission map estimation module performs a connection operation between the background scattering estimation and the input image. The operation includes 3 sets of dilated convolutional kernels with a size of 5×5, one set of convolutional kernels with a size of 3×3, and one set of convolutional kernels with a size of 1×1. The number of each set of dilated convolutional kernels is 8, and the number of the last 2 sets of ordinary convolutional kernels is 3. Dilated convolution can increase the receptive field by filling additional 0s in the convolutional kernel without changing the number of parameters. A larger receptive field can help the network use more image information to estimate the direct transmission parameters. To avoid the grid effect brought by dilated convolution, different dilation rates are selected for continuous dilated convolution operations.

[0028] Step 3: Reconstruct the underwater image based on the above background light estimation B and transmission map estimation A, specifically:

[0029] In the underwater environment, due to the attenuation and absorption of light by the water medium, the underwater environment is often a low-light environment, especially in the night environment. The underwater low-light enhancement task aims to correct the image contrast and restore the details of the dark areas of the image hidden underwater. Current research has proposed a wide range of underwater low-light enhancement methods. Among them, the method based on the Jaffe-McGlamery model has received wide attention. The Jaffe-McGlamery theory believes that the underwater imaging model is the superposition of direct light and reflected light

[0030] I c (x) = J c (x)t(x) - B(1 - t(x))c ∈ {r, g, b} (2)

[0031] Where x is the pixel coordinate of the image, c is the red, green, and blue channels of the image, r, g, b represent red, green, and blue respectively, I(x) is the underwater image collected, J(x) is the underwater image without attenuation and scattering by the water medium, that is, the enhanced image we want to obtain. t(x) is the transmission map of the underwater image, and B is the background light of the underwater image.

[0032]

[0033] Where β represents the attenuation coefficient of the three channels in the same water medium. Usually, β r > β g > β b , because the absorption of long - wave light by water is greater than that of short - wave light. Therefore, the attenuation of the three light waves of red, green, and blue in water decreases in turn.

[0034] Substitute the estimated background light B and the estimated transmission map t obtained in Steps 1 and 2 into formula (2) to obtain the reconstructed image I0.

[0035] Step 4: Enhance the details of the local and global parts of the image through multi - scale Gaussian blur. Specifically:

[0036] Enhance the details of the local and global parts of the image through multi - scale Gaussian blur, solve the detailed texture of the image, and enhance it on this basis.

[0037]

[0038] L1 = I * G1 L2 = I * G2 L3 = I * G3 (5)

[0039] B1 = I - L3 B2 = I - L2 B3 = I - L1 (6)

[0040] Where G represents the Gaussian function, x is the pixel offset from the center point, σ is the standard deviation of the Gaussian function, which controls the degree of blur, π is the pi; I is the reconstructed image, G1, G2, G3 are Gaussian blur functions, and their Gaussian kernels are 1, 5, and 10 respectively; L1, L2, L3 are the images blurred by different Gaussian kernels; B1, B2, B3 are the image texture maps of different degrees. Among them, B1 contains the most high - frequency texture information, B2 is the second, and B3 is the least, reflecting the different - scale - blur - based differential texture of the image detail retention degree.

[0041] Step 5, set the weights ω1, ω2, ω3 of the texture maps, and output the enhanced image I * , specifically as follows:

[0042] I * = I + ω1B1 + ω2B2 + ω2B3

[0043] I * represents the enhanced output image, and ω1, ω2, ω3 are the weights of the three texture maps respectively. The weights are set as ω1 = ω2 = ω3 = 1 / 3.

[0044] The image materials for subjective experiment comparison are respectively selected from natural images in different environments of the UIEB dataset and the EVUP dataset.

[0045] The present invention designs a CNN model based on physical model constraints. Existing methods aim to learn better solutions from training data, but lack physical model constraints, and it cannot be guaranteed whether the solutions obtained satisfy the physical model. In contrast, we use the underwater imaging model as the constraint of the CNN model, effectively reducing the ill-posedness of the image enhancement problem, providing model constraint conditions, so that the resulting image has higher consistency with the observed image.

[0046] The present invention uses a high-quality underwater image dataset enhanced by deep learning methods as the training set, combines the advantages of physical models and deep learning, and fits the physical model parameters with this dataset, making the parameters as close as possible to the model parameters required for real underwater image reconstruction. Compared with other deep learning methods, this method has a simple structure and achieves better results in subjective and objective analyses.

[0047] The present invention adds an image detail enhancement model on the basis of the above model, performs multi-scale optimization on the results on the basis of the original model results, improves the contrast and details of the image, reduces the blur problem existing in underwater images, and effectively enhances real-world underwater images.

[0048] Next, in combination with the embodiments, the underwater image enhancement method based on the convolutional network constrained by the underwater imaging physical model proposed by the present invention is compared with the conventional underwater image enhancement methods based on the dark channel prior and the MIP method.

[0049] Embodiment

[0050] First, the details of the experimental setup are introduced. Through visual subjective evaluation and quantitative evaluation, the superiority of the method of the present invention is verified. This embodiment is carried out in the Windows 11 system, and the algorithm of the present invention is implemented using the PyTorch deep learning framework. The batch size used in deep learning training is 1, the learning rate is 0.001, and the number of training epochs is 1000. The image source of the experiment is the UIQS dataset. There are 890 pairs of matching images in the dataset for training, and our method is tested on the UIQS dataset. The UIQS dataset is a dataset covering various underwater detection challenges. The labeled target categories in this dataset include: fish, diver, starfish, coral, turtle, sea urchin, sea cucumber, scallop, squid, jellyfish, a total of 10 categories. Among the evaluation metrics, we adopt common image quality evaluation metrics, full-reference measures and non-reference metrics, including average gradient (AG), underwater image quality metric (UIQM, UCIQE), structural similarity index measure (SSIM and PSNR). The algorithm of the present invention is compared and verified with 8 methods, namely DCP, GBdehazingRCorrection, IBLA, LowComplexityDCP, MIP, RoWS, UDCP, ULAP.

[0051] Figure 4Shows 8 underwater images under different degradation conditions. In Picture1 - 5, we can see that the original images have an overall green background. Based on the DCP and MIP methods, neither can well achieve the compensation of the color channels of the images, and the overall green background is not well improved. The GBdehazingRCorrection method based on red channel compensation has an obvious over - compensation problem, making the overall background of the image red. The UDCP and the algorithm proposed in the present invention well compensate the color channels, better eliminate the influence of the green background of the image, and at the same time clearly restore the features of the objects in the underwater image, facilitating the visible recognition of the objects in the underwater image by the human eye. In Picture6, it can be seen from the original image that it is a sea turtle, but the overall background of the image is blue, and the image itself is blurred, and the texture features of the sea turtle are not obvious. Based on the DCP and MIP methods, the color channel compensation effect of the underwater image with a blue background is not good, and the blurriness of the image is not well eliminated. The image processed by the GBdehazingRCorrection method is overall darker, which is often due to estimating a larger value in the process of background light estimation, resulting in the overall image being darker during the restoration process. The image obtained by the LowComplexityDCP method is brighter and the overall image tends to be bright light blue, resulting in serious masking of the features of the sea turtle itself, and the texture features of the sea turtle are even less clear. The UDCP has an obvious problem of excessive enhancement of details, and the local detailed textures are not coherent. Compared with other algorithms, the method proposed in the present invention not only better restores the texture features of the sea turtle itself but also achieves a better color compensation effect. Picture7 is an underwater image taken at close range, overall light blue. Based on the DCP method, the color compensation is not well processed and is not much different from the original image. The MIP method has a good processing of the original image and better eliminates the influence of the blue background. The GBdehazingRCorrection and ULAP obviously over - compensate the color channels of the image, resulting in the image showing an abnormal red color. The LowComplexityDCP method makes the image darker and the details unclear. Our method not only has a good restoration of the image color but also has a better contrast than the MIP, and the colors of the image objects have a better saturation. In Picture8, the DCP method eliminates the influence of blurriness to a certain extent, but the effect of color restoration is not good. The MIP method makes the image darker and masks the texture features of the people in the image. The IBLA method has a good restoration of the figure, but the plants under the people are still darker. Compared with the comparison methods, the method proposed in the present invention makes the texture features of the people and plants in the image better appear, and is more visually appealing, but there is a certain degree of over - exposure in the background.In the underwater images presented in Picture 1-8 under different lighting conditions, the results of these ten methods show that under different lighting conditions, the method of the present invention can well restore the color of the image, effectively eliminate the blue-green bias in the underwater image, and restore the detailed textures of animals and plants better.

[0052] In this embodiment, full-reference and non-reference metrics are used to evaluate the quality of the restored images. The statistical results are shown in Tables 1-5.

[0053] Table 1 SSIM data analysis

[0054] Method SSIM DCP 0.856 GBdehazingRCorrection -0.051 IBLA 0.669 LowComplexityDCP 0.335 MIP 0.712 RoWS 0.941 UDCP 0.511 ULAP 0.684 Ours 0.597

[0055] Table 2 PSNR data analysis

[0056]

[0057]

[0058] Table 3 Average gradient

[0059]

[0060] Table 4 UIQM data analysis

[0061] Method UIQM UICM UISM UICONM DCP 0.418 5.053 1.035 -0.044 GBdehazingRCorrection 0.162 4.847 0.434 -0.151 IBLA 0.290 12.647 0.252 -0.208 LowComplexityDCP 1.265 6.340 0.335 -0.104 MIP 0.154 9.224 1.009 -0.182 RoWS 0.112 4.068 0.351 -0.156 UDCP 0.091 7.885 -0.427 -0.007 ULAP 0.158 4.001 0.375 -0.096 Ours 0.307 9.760 0.542 -0.189

[0062] Table 5 UCIQE data analysis

[0063] Method UCIQE DCP 60.361 GBdehazingRCorrection 61.361 IBLA 61.068 LowComplexityDCP 65.812 MIP 56.187 RoWS 54.602 UDCP 66.787 ULAP 59.817 Ours 55.917

[0064] Due to different underwater environments, the reference images used for evaluation are selected from existing open-source underwater databases. The full-reference metric is used to evaluate the similarity between our proposed method and nine methods. The obtained results imply that the restored images generated by our proposed method are different from those generated by the nine compared methods. To evaluate the quality of restored images in a harsh underwater environment, 136 challenging underwater images were used in the experiment. Compared with ordinary underwater images, challenging images suffer more degradation, such as lack of details, high backscatter scenes, and insufficient lighting conditions. The present invention conducted a full-reference evaluation using SSIM and PSNR metrics, which are used to represent the similarity between the restored image and the reference image, and the results are shown in Tables 1-2. In addition, as shown in Tables 3-5, our proposed method achieved relatively high values of the average gradient (AG), contrast metric (UIQM), underwater image chromaticity metric (UICM), and underwater image sharpness metric (UISM) of underwater images. In the UCIQE of the three datasets, our method achieved a relatively high score. As Figure 4As shown, we selected 8 challenging underwater images from the data. The images restored by the proposed method are better than the 8 compared methods. The proposed method achieves better performance than other haze removal algorithms in harsh environments.

[0065] Full-reference evaluation: Full-reference evaluation can provide as much realistic feedback as possible, using real ground truth reference images to evaluate the performance of different methods. We selected two classic metrics (PSNR and SSIM) for full-reference evaluation. Higher PSNR and SSIM scores indicate restored images that are closer to the reference results. The average scores on PSNR and SSIM of the SUID dataset are listed in Tables 1 - 2. Among the 9 compared methods, the RoWS method has the best scores on PSNR and SSIM. The method proposed in this invention is slightly inferior. The main reason is that the algorithm proposed in this invention estimates the image depth and uses it as one of the bases for image restoration, resulting in a certain degree of white overexposure in the restored image compared to the original image. While the GBdehazingRCorrection method does not have the process of depth estimation, so the restored image does not effectively remove the blue-green color, but does not introduce excessive white overexposure. The PSNR and SSIM methods are more sensitive to the error between images, resulting in slightly lower results for the method of this invention.

[0066] Non-reference evaluation: We conducted non-reference evaluations on the RUOD and EUVP databases. We selected three classic non-reference metrics (AG, UIQM, UCIQE) for evaluation. As can be seen in Table 3, the proposed method in this invention obtains the highest average gradient value, indicating that the resulting images of the proposed method in this invention have more obvious textures compared to other methods. However, in UIQM and UCIQE, as shown in Tables 4 - 5, the method of this invention does not obtain the highest scores. This is because although UIQM and UCIQE are evaluation metrics designed based on human vision, the evaluation metrics of images do not always match human visual perception in different situations, especially in the underwater environment. This metric can only be used as a reference and cannot be used as a judgment standard. Because although the NewOpticalModel method obtains the best scores, in subjective evaluation, its resulting images have obvious problems, such as the image not being able to effectively remove the blue-green color cast, excessive red compensation, and insufficient brightness. Therefore, we believe that there is a gap between the quantitative scores of non-reference metrics and subjective visual quality. The image quality evaluation metrics designed for underwater images have limitations in some cases.

Claims

1. An underwater image enhancement method based on a convolutional network constrained by an underwater imaging physical model, characterized in that It includes the following steps: Step 1: Construct a background light estimation module to obtain the background light estimation of the input image; Step 2: Construct a transmission map estimation module to obtain the transmission map estimation; Step 3: Use an underwater physical model to restore the underwater image; Step 4: Solve the detailed texture map of the underwater image; Step 5: Set the texture map weight and output the enhanced image.

2. The underwater image enhancement method of the convolutional network based on the constraint of the underwater imaging physical model according to claim 1, wherein In Step 1, the background light estimation module is used to calculate the background light estimation of the input underwater image. Specifically: The background light estimation module consists of 2 groups of convolutional kernels with a size of 5×5, 2 global average pooling layers, 2 groups of convolutional kernels with a size of 3×3, and 2 groups of convolutional kernels with a size of 1×1. The number of each group of convolutional kernels is 3; the activation function for the convolutional operation selects the Mish activation function; f(x) = x * tanh(ln(1 + e x )) (1) where x represents an input variable and is the output of a certain layer in the neural network; e x is the natural exponential function, representing the exponential with the natural constant e as the base; ln(1 + e x ) is the Softplus function, which is used to smoothly map the input into the positive range; tanh(·) represents the hyperbolic tangent function, and its value range is (–1, 1).

3. The underwater image enhancement method of the convolutional network based on the constraint of the underwater imaging physical model according to claim 1, characterized in that, In Step 2, the transmission map estimation module is used to calculate the transmission map estimation of the input underwater image. Specifically: The transmission map estimation module performs a connection operation between the background scattering estimation and the input image. The operation includes 3 groups of dilated convolutional kernels with a size of 5×5, one group of convolutional kernels with a size of 3×3, and one group of convolutional kernels with a size of 1×1. The number of each group of dilated convolutional kernels is 8, and the number of the last 2 groups of ordinary convolutional kernels is 3; the dilated convolution can increase the receptive field by filling extra 0s in the convolutional kernel without changing the number of parameters, and different dilation rates are selected for continuous dilated convolutional operations.

4. The underwater image enhancement method of the convolutional network based on the constraint of the underwater imaging physical model according to claim 1, wherein Based on the background light estimation and the transmission map estimation, the underwater image is reconstructed. Specifically: According to the Jaffe-McGlamery theory, the underwater imaging model is the superposition of direct light and reflected light I c J(x) = c J(x)t(x) - B(1 - t(x))c ∈ {r, g, b} (2) where x is the pixel coordinate of the image, c is the red, green, and blue channels of the image, I(x) is the acquired underwater image, J(x) is the underwater image without being attenuated and scattered by the water medium; t(x) is the transmission map of the underwater image, and B is the background light of the underwater image; where β represents the attenuation coefficient of the three channels in the same water medium, β r > β g > β b , since water absorbs more of the long wavelengths of light than the short wavelengths, the attenuation of the red, green, and blue light waves underwater decreases in turn; Substitute the background light estimation B and the transmission map estimation t obtained in Step 1 and Step 2 into formula (2) to obtain the reconstructed image I0.

5. The underwater image enhancement method of the convolutional network based on the constraint of the underwater imaging physical model according to claim 1, characterized in that, In Step 4, the local and global details of the image are enhanced by means of multi-scale Gaussian blur. Specifically: The local and global details of the image are enhanced by means of multi-scale Gaussian blur, the detailed texture of the image is solved, and it is enhanced on this basis; where G represents the Gaussian function, x is the pixel offset from the center point, σ is the standard deviation of the Gaussian function, which controls the degree of blur, and π is the circumference ratio; L1 = I * G1 L2 = I * G2 L3 = I * G3 (5) B1 = I - L3 B2 = I - L2 B3 = I - L1 (6) where I is the reconstructed image, G1, G2, G3 are Gaussian blur functions, and their Gaussian kernels are 1, 5, and 10 respectively; L1, L2, L3 are the images blurred by different Gaussian kernels; B1, B2, B3 are the image texture maps with different degrees, where B1 contains the most high-frequency texture information, B2 is the second, and B3 is the least, reflecting the differential texture of the image detail retention degree under different scale blurs.

6. The underwater image enhancement method of the convolutional network based on the constraint of the underwater imaging physical model according to claim 5, characterized in that, Set the weights ω1, ω2, ω3 of the texture map and output the enhanced image I * , specifically as follows: I * = I + ω1B1 + ω2B2 + ω2B3 I * represents the enhanced output image, where ω1, ω2, and ω3 are the weights of the three texture maps respectively.

7. The underwater image enhancement method of the convolutional network based on the constraint of the underwater imaging physical model according to claim 6, characterized in that ω1 = ω2 = ω3 = 1 / 3.

8. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images