A low-light underwater image enhancement method and system

Abstract

This invention discloses a method and system for enhancing low-light underwater images, including constructing a low-light underwater image dataset, constructing a low-light underwater image enhancement network model, training the constructed low-light underwater image enhancement network model based on the low-light underwater image dataset, and finally using the trained low-light underwater image enhancement network model to predict low-light underwater image test samples to obtain high-quality underwater images, which are then evaluated using evaluation metrics. The solution of this invention employs a Retienx decomposition module based on deep learning, which effectively alleviates the specific scene limitations of pure physical algorithms. Utilizing a decoder feature module and a global illumination feature module, it promotes information interaction between encoding layer features and decoding layer features at the global level, thereby greatly improving the model's restoration capability. Extensive experiments have demonstrated the effectiveness of this method, and experiments on real low-light underwater image datasets show that the proposed method has good performance.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement, specifically relating to a method and system for enhancing underwater images in low light. Background Technology

[0002] With increasing pressure on land space and resources, marine resources are gradually becoming a new global focus, and underwater images are crucial for observing the underwater environment. However, the uneven and wavelength-dependent attenuation of light in water leads to significant color deviations in underwater images. Furthermore, water particles reflect light, blurring details and affecting image quality. These complex underwater environments result in color distortion and high noise levels. While underwater exploration technology has advanced to greater depths, low visibility in deep-sea operations causes underwater images to suffer from color shifts, low brightness, and severe noise. Current deep-learning-based underwater image enhancement algorithms struggle to effectively improve the visual quality of low-light underwater images due to complex degradation. Although low-light image enhancement algorithms in natural scenes can effectively correct illumination and reduce noise, they are less effective at addressing color distortion caused by underwater imaging processes. In summary, while current underwater image enhancement technologies and low-light image enhancement algorithms can solve some of the problems of low-light underwater images, they still cannot meet the effect requirements of low-light underwater image enhancement. Summary of the Invention

[0003] To address the aforementioned problems, the present invention aims to provide a method and system for enhancing underwater images in low light conditions, which focuses on both illumination restoration and color correction of underwater images.

[0004] The specific technical solution for achieving the objective of this invention is as follows:

[0005] A method for enhancing underwater images in low light includes the following steps:

[0006] Step 1: Construct a low-light underwater image dataset;

[0007] Step 2: Construct a network model for enhancing low-light underwater images;

[0008] Step 3: Train the constructed low-light underwater image enhancement network model based on the low-light underwater image dataset;

[0009] Step 4: Use the trained low-light underwater image enhancement network model to predict the test samples of low-light underwater images to obtain high-quality underwater images, and use evaluation metrics for evaluation.

[0010] The present invention also provides a low-light underwater image enhancement model, comprising the following modules:

[0011] Model building module: used to build low-light underwater image datasets and low-light underwater image enhancement network models;

[0012] Training module: Used to train the constructed low-light underwater image enhancement network model based on the low-light underwater image dataset;

[0013] Prediction and Evaluation Module: This module uses a trained low-light underwater image enhancement network model to predict test samples of low-light underwater images, obtain high-quality underwater images, and evaluate them using evaluation metrics.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0015] Compared with existing technologies, the method of the present invention recovers a high-quality underwater image through a depth Retinex decomposition module and multiple encoding and decoding modules including illumination light attention and global illumination feature modules;

[0016] Specifically, the Retinex decomposition module is first applied to the input image to obtain the reflected light component and the illumination light component. The Retinex decomposition module based on deep learning can effectively alleviate the specific scene limitations brought about by pure physical algorithms.

[0017] Subsequently, this invention performs convolution operations on the illumination light component and fuses and stitches the input original low-light underwater image to obtain an initial feature map. Here, the stitching and fusion operations supplement the lost details of the features. Next, this invention applies multiple feature encoding and feature decoding modules to the initial features and fuses the reflected light component to obtain a depth feature map. The encoding feature module and the decoder feature module can efficiently capture long-distance dependencies between features by utilizing illumination light attention. The multi-scale fusion module can efficiently fuse feature context information at different scales. In addition, the decoder feature module can promote information interaction between the encoding layer features and the decoding layer features at the global level by utilizing the global illumination feature module, thereby greatly improving the model's repair capability.

[0018] This invention trains a network by minimizing the mean squared error loss and structural similarity index loss between high-quality underwater images and their corresponding real-world underwater image standards, as well as the loss between reflected light components and illumination light components, to improve the network's inpainting ability. Extensive experiments demonstrate the effectiveness of this method, and experiments on real-world low-light underwater image datasets show that the proposed method exhibits good performance.

[0019] The present invention will be further described below with reference to specific embodiments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the low-light underwater image enhancement method of the present invention.

[0021] Figure 2 This is a diagram of the low-light underwater image enhancement network framework in an embodiment of the present invention. Detailed Implementation

[0022] Example

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0025] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0026] Combination Figure 1 A low-light underwater image enhancement method includes the following steps:

[0027] Step 1: Construct a low-light underwater image dataset S = {(x1,y1),(x2,y2),...,(x...} L ,y L )};

[0028] Where, x iRepresents a low-light underwater image, y i Indicates with x i The corresponding high-quality real underwater images, i = 1, ..., L, where L represents the number of samples in the dataset;

[0029] The minimum-maximum normalization method was used to control the data range in the low-light underwater image dataset to be between 0 and 1.

[0030] Step 2, Combining Figure 2 A low-light underwater image enhancement network model was constructed, including a deep Retinex decomposition module, multiple cascaded feature encoders and feature decoders, and a feature fusion module.

[0031] In this embodiment, the feature encoder includes three modules and the feature decoder module includes two modules.

[0032] The process of low-light underwater image enhancement network model performing low-light underwater image enhancement is as follows:

[0033] First, input a low-resolution, low-light image I. l The light source is decomposed based on the deep Retinex decomposition module, which includes stitching and dimensionality increase operations and multiple convolution operations. First, the input low-light underwater image I... l The features are concatenated to increase the dimensionality, resulting in initial decomposed features. Then, 9x9 convolution operations are used for feature enhancement, followed by five sets of 3x3 convolutions and ReLU activation functions for further feature enhancement and extraction. Finally, the decomposed reflected light component R and illumination light component L are obtained, specifically:

[0034] For the input low-light underwater image I l The features are concatenated to increase dimensionality, resulting in initial decomposed features. Then, 9x9 convolution operations are used for feature enhancement, followed by five sets of 3x3 convolutions and ReLU activation functions for further feature enhancement and extraction. Finally, the obtained features are segmented into reflected light component R and illumination light component L.

[0035]

[0036] Among them, I l The input is a low-light underwater image. [·] concatenates the two vectors along the channel dimension. Conv 9*9 and Conv 3*3 It is a convolution operation, ReLU is an activation function, Chunk is a feature segmentation operation, R is the reflection component, and L is the illumination component.

[0037] After feature extraction of the illumination light component L, the input image I is fused. lAn initial feature map F is obtained. After processing the initial feature map F through three cascaded feature encoders and two feature decoders, the reflected light component R is fused to obtain the enhanced depth feature map F. d :

[0038] The illumination light component L is subjected to 3*3 convolution feature extraction, and the features of the input original underwater image are fused to generate an initial feature map F:

[0039] F = Conv 3*3 (L)⊙I l +I l

[0040] Where L is the illumination light component, I l It is the input low-light underwater image, Conv 3*3 The convolution operation F is the initial feature map;

[0041] Each of the multiple cascaded feature encoders includes a 4*4 convolutional unit and an illumination attention unit. The illumination attention unit includes an illumination attention layer and a multi-scale feature fusion layer to enhance the input features.

[0042] The processing procedure for each feature encoder is as follows:

[0043] F L =Conv 4*4 (F t ),

[0044] F c =RA(LN(F) L ),F R )+F L

[0045] F e =MFFB(LN(F c ))+F c

[0046] Among them, F t The input feature map is the illumination light component L, and the input features of each subsequent feature coding layer are the output features of the previous feature coding layer. 4*4 It is a convolution operation, F L It is the input feature of the illumination attention unit, F R The input reflected light features are obtained by dimensionality reduction of the reflected light component R. LN is the layer normalization operation, and F... c It is an intermediate feature, F e These are the output features. RA represents the illumination light attention layer processing process, and MFFB represents the multi-scale feature fusion layer processing process.

[0047] The remaining two feature encoders sequentially enhance the features output by the feature module of the previous encoder to obtain the final encoded feature output.

[0048] Among them, the RA illumination attention layer mainly enhances the long-distance dependencies between features. First, it processes the input features using F... LN After row dimension reshaping, a 1D convolution operation is used to aggregate pixel-level cross-channel context to obtain query F. Q Key F k Sum F V The characteristics of the reflected light F of the input R The system also performs dimension reshaping and 1D convolution operations, multiplies the resulting feature vectors and values ​​to obtain the latest values, and finally applies the self-attention rule to calculate the output feature F. a The processing procedure is as follows:

[0049]

[0050] F a =Conv 1*1 (F V ⊙F R ·Softmax(F Q ·F K ))+F LN

[0051] Among them, F LN It is F L F′ is obtained by applying the input features after layer normalization. LN It is F LN After row dimension reshaping, the feature F′ R It is F R The input reflected light features after row dimension reshaping, Conv 1*1 It is a convolution operation, F Q It's a query, F K It is the key, F V It is a value, F R It is the characteristic of reflected light F R The value of F LN It is the input feature, F a It is the output feature;

[0052] The MFFB multi-scale feature fusion layer mainly performs interactive fusion of the enhanced features. First, it processes the input feature F... CL Perform convolution operations, then apply depthwise convolutions of 3x3 and 5x5 size and ReLU activation functions to the features respectively. Next, concatenate these features, and apply depthwise convolutions of 3x3 and 5x5 size and ReLU activation functions to the features again. Finally, concatenate and apply convolution operations to obtain the output feature F.M The specific processing procedure is as follows:

[0053] F n =Conv 1*1 (F CL )

[0054]

[0055] F M =Conv 1*1 ([F′ x ,F′ y ])

[0056] Among them, F CL It is F C After layer normalization of the input features, Conv 1*1 It's a convolution operation, DW 3*3 and DW 5*5 This is a depthwise convolution operation, Chunk is a feature segmentation operation, [·] concatenates two vectors along the channel dimension, F M It is the output feature;

[0057] The two cascaded feature decoders each include a 2x2 depthwise convolutional layer, a feature fusion layer, a 1-dimensional convolutional layer, a global illumination feature layer, and an illumination attention layer. First, for the input feature F... e A 2x2 depthwise convolution operation is performed, and the fusion module combines the features F′ from the encoding layer through a concatenation operation. e The fusion process is then performed, followed by 1-dimensional convolution and global illumination feature module operations on the fused features. Finally, an illumination attention module is used to process the features and reflected light features F. R Attention enhancement is performed, and the final output feature F is obtained. d The remaining decoders sequentially enhance the features output by the feature module of the previous decoder to obtain the final decoder feature output.

[0058] The processing procedure for each feature decoder is as follows:

[0059] F L =FA(Conv 1*1 ([DW 2*2 (F e ),F′ e ])

[0060] F C =RA(LN(F) L ),F R )+F L

[0061] F d =MFFB(LN(FC ))+F c

[0062] Among them, F e The input features are the initial feature decoding layer's input features, which are the output features of the last feature encoding layer. Subsequent feature decoding layers' input features are the output features of the previous feature decoding layer. (DW) 2*2 It is a depthwise convolution operation, F′ e It is the output feature of the symmetric hierarchical feature encoding layer, [·] indicates that two vectors are concatenated along the channel dimension, Conv 1*1 It is a convolution operation, FA represents the global illumination feature layer processing, F L It is the input feature of the illumination attention module, F R The input reflected light features are obtained by dimensionality reduction of the reflected light component R. LN is the layer normalization operation, RA represents the illumination light attention processing process, MFFB represents the multi-scale feature fusion layer processing process, and F... C It is an intermediate feature, F d It is the output feature;

[0063] In this process, the global illumination feature layer enhances the intermediate features of the fused coding layer by first processing the input feature F. cn Adaptive average pooling convolutions of different dimensions are performed. First-dimensional depthwise convolutions are applied to the features in dimensions H and W, followed by cross products. A 3x3 convolution is then applied to the cross product to obtain intermediate features. First-dimensional convolutions are applied to the features in dimension C, followed by Sigmoid activation. Then, a dot product is performed with the intermediate features, followed by a first-dimensional convolution and Sigmoid activation function. Finally, a dot product is performed with the initial input features to obtain the final output feature F. f The processing procedure is as follows:

[0064] F H =DW 1*1 (AAP H (F cn )),F W =DW 1*1 (AAP W (F cn )),F C

[0065] =Sig(Conv 1*1 (AAP C (F cn )))

[0066] F f =Sig(CNN) 1*1 (Conv 3*3 (F H ·FW )⊙F C ))⊙F cn

[0067] Among them, F cn It is the input feature map, AAP is adaptive average pooling convolution, and DW 1*1 It is a depthwise convolution operation, Conv 1*1 and Conv 3*3 It's a convolution operation; Sig is the sigmoid activation function, F... f It is the output feature;

[0068] The illumination attention processing process in the feature decoder is the same as the illumination attention layer processing process in the feature encoder.

[0069] The multi-scale feature fusion layer processing procedure in the feature decoder is the same as that in the feature encoder.

[0070] Finally, based on the feature fusion module, the deep feature map F is processed. d The feature fusion module includes a 3x3 convolution and residual operations. First, a 3x3 convolution is applied to the depth feature map F. d Dimensionality reduction is performed, and then the initial feature map F is fused to obtain a high-quality, clear underwater image enhancement image I. h :

[0071] I h =Conv 3*3 (F d )+F

[0072] Among them, F d It is the acquired depth feature map, Conv 3*3 It is a convolution operation, F is the initial feature map, and I is... h It is a high-quality, clear underwater image enhancement image.

[0073] Step 3: Train the constructed low-light underwater image enhancement network model based on the low-light underwater image dataset;

[0074] During training, the training objective is to minimize the mean squared error loss and structural similarity index loss between high-quality underwater images and their corresponding real underwater image standards, as well as the loss between reflected light components and illumination light components.

[0075] The objective function for underwater image enhancement is represented by a loss function, and the training of a network model for low-light underwater images is completed.

[0076]

[0077] Lr =||I h -I gt ||2

[0078] L ssim =1-SSIM(I h ,I gt )

[0079]

[0080] Among them, I h I represents the underwater image predicted by the network. gt Indicates the underwater image I predicted by the network. h The corresponding real underwater image, ||·||² represents the mean squared error loss, SSIM is the structural similarity index loss, R h L represents the reflected light component of the underwater image predicted by the network. h R represents the underwater image illumination light component predicted by the network. gt L represents the reflected light component of a real underwater image. gt The illumination light component represents the actual underwater image, and ||·||1 represents the absolute error loss.

[0081] Step 4: Use the trained low-light underwater image enhancement network model to predict the test samples of low-light underwater images to obtain high-quality underwater images, and use evaluation metrics for evaluation.

[0082] The evaluation metrics include peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM).

[0083]

[0084]

[0085] Among them, I i h and I i gt These represent the high-quality underwater images I belonging to the model output. h The pixels and I h Corresponding real underwater image I gt pixels; MAX I This represents the maximum value of the color of an image point, typically 255; n represents the number of samples in the training set; and i represents the current sample.

[0086] c1 = (k1L) 2

[0087] c2=(k2L) 2

[0088]

[0089] Where, μ Ih It is a high-quality underwater image output by the model. h The average value, Is with I h Corresponding real underwater image I gt The average value, isI h variance isI gt The variance; L is the dynamic range of pixel values, and k1 and k2 are both preset hyperparameters, which are set to: k1 = 0.01, k2 = 0.03. Y represents pred Standard deviation and Y gt Standard deviation The product of c1 and c2, where c1 and c2 are smoothing parameters, is used to compare the brightness, contrast and structural characteristics of images.

[0090] The test data was enhanced to obtain high-quality underwater images, which were then evaluated using the aforementioned customized metrics. The effectiveness of the proposed method was demonstrated through extensive experiments and applied to low-light underwater images.

[0091] This embodiment uses the above two indicators to evaluate the completion results. Two representative algorithms in the field of underwater image are selected: Ucolor and Ushape. The indicator results are shown in Table 1. The method of this invention (Ours) has achieved the best results in both PSNR and SSIM, indicating that the enhancement results of this invention are closer to the real results.

[0092] Table 1 Results of image enhancement indices for underwater images under low light conditions

[0093]

[0094] The present invention also provides a low-light underwater image enhancement model, comprising the following modules:

[0095] Model building module: used to build low-light underwater image datasets and low-light underwater image enhancement network models;

[0096] Training module: Used to train the constructed low-light underwater image enhancement network model based on the low-light underwater image dataset;

[0097] Prediction and Evaluation Module: This module uses a trained low-light underwater image enhancement network model to predict test samples of low-light underwater images, obtain high-quality underwater images, and evaluate them using evaluation metrics.

[0098] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for enhancing underwater images in low light, characterized in that, Includes the following steps: Step 1: Construct a low-light underwater image dataset; Step 2: Construct a low-light underwater image enhancement network model, including: a deep Retinex decomposition module, multiple cascaded feature encoders and feature decoders, and a feature fusion module; The process of low-light underwater image enhancement network model performing low-light underwater image enhancement is as follows: First, input a low-resolution, low-light image. The light source is decomposed based on the depth Retinex decomposition module to obtain the decomposed reflected light components. and illumination light components : For input low-light underwater images The features are concatenated to increase their dimensionality, resulting in initial decomposed features. Then, 9x9 convolution operations are used for feature enhancement, followed by five sets of 3x3 convolutions and ReLU activation functions for further feature enhancement and extraction. Finally, the obtained features are segmented to form reflected light components. and illumination light components : ； in, It is an input low-light underwater image. Connect the two vectors along the channel dimension. and It's a convolution operation. It is an activation function. It involves segmenting features. It is the reflection component. It is the component of illumination light; For the illumination light component After feature extraction, the input image is fused. Obtain the initial feature map For the initial feature map After processing by multiple cascaded feature encoders and feature decoders, the reflected light components are fused. The enhanced depth feature map is obtained. ; Finally, the feature fusion module is used to process the deep feature map. By fusing the images, high-quality underwater images can be obtained. ; Step 3: Train the constructed low-light underwater image enhancement network model based on the low-light underwater image dataset; Step 4: Use the trained low-light underwater image enhancement network model to predict the test samples of low-light underwater images to obtain high-quality underwater images, and use evaluation metrics for evaluation.

2. The low-light underwater image enhancement method according to claim 1, characterized in that, The low-light underwater image dataset in step 1 is... ; in, This represents a low-light underwater image. Indicates and Corresponding high-quality real underwater images, , Indicates the number of samples in the dataset; The minimum-maximum normalization method was used to control the data range in the low-light underwater image dataset to be between 0 and 1.

3. The low-light underwater image enhancement method according to claim 1, characterized in that, The enhanced depth feature map The process is as follows: For the illumination light component 3x3 convolution feature extraction is performed and fused with the features of the input original underwater image to generate an initial feature map. : ； in, It is the component of illumination light. It is an input low-light underwater image. It is a convolution operation It is the initial feature map; Each of the multiple cascaded feature encoders includes a 4*4 convolutional unit and an illumination attention unit. The illumination attention unit includes an illumination attention layer and a multi-scale feature fusion layer to enhance the input features. The processing procedure for each feature encoder is as follows: ； ； ； in, The input feature map is the illumination light component, which is the input feature of the initial feature encoding layer. The input features of each subsequent feature coding layer become the output features of the previous feature coding layer. It's a convolution operation. These are the input features of the illumination attention unit. It is the characteristic of the input reflected light, composed of the reflected light components. The dimensionality reduction process is performed to obtain the result. It is a layer normalization operation. It is an intermediate feature. It is the output feature. This indicates the illumination attention layer processing process. This represents the multi-scale feature fusion layer processing procedure; Each of the multiple cascaded feature decoders includes a 2*2 depth convolutional layer, a feature fusion layer, a 1-dimensional convolutional layer, a global illumination feature layer, and an illumination attention layer. The processing procedure for each feature decoder is as follows: ； ； ； in, The input features are the input features of the initial feature decoding layer, which are the output features of the last feature encoding layer. The input features of subsequent feature decoding layers are the output features of the previous feature decoding layer. It is a depthwise convolution operation. It is the output feature of the symmetric hierarchical feature encoding layer. This indicates that two vectors are concatenated along the channel dimension. It's a convolution operation. This represents the global illumination feature layer processing procedure. These are the input features of the illumination attention layer. It is the characteristic of the input reflected light, composed of the reflected light components. The dimensionality reduction process is performed to obtain the result. It is a layer normalization operation. This indicates the illumination attention layer processing process. This represents the multi-scale feature fusion layer processing procedure. It is an intermediate feature. It is the output feature.

4. The low-light underwater image enhancement method according to claim 3, characterized in that, The illumination attention layer processing procedure in the feature encoder is as follows: ； ； in, yes Input features after layer normalization. yes Features after row dimension reshaping yes Input reflected light features after row dimension reshaping It's a convolution operation. It's a query. It is a key. It is worth it. It is a characteristic of reflected light The value, These are input features. It is the output feature; The multi-scale feature fusion layer processing procedure in the feature encoder is as follows: ； ； ； ； in, yes Input features after layer normalization It's a convolution operation. and It is a depthwise convolution operation. It is a feature segmentation operation. Connect the two vectors along the channel dimension. It is the output feature; The global illumination feature layer processing procedure in the feature decoder is as follows: ； ； in, It is the input feature map. It is adaptive average pooling convolution. It is a depthwise convolution operation. and It's a convolution operation. It is the Sigmoid activation function. It is the output feature; The illumination attention processing process in the feature decoder is the same as the illumination attention layer processing process in the feature encoder. The multi-scale feature fusion layer processing procedure in the feature decoder is the same as that in the feature encoder.

5. The low-light underwater image enhancement method according to claim 1, characterized in that, The underwater images The acquisition process is as follows: ； in, It is the acquired depth feature map. It's a convolution operation. This is the initial feature map.

6. The low-light underwater image enhancement method according to claim 1, characterized in that, In step 3, the training objective of the low-light underwater image enhancement network model is to minimize the mean square error loss and structural similarity index loss between the high-quality underwater image and its corresponding real underwater image standard, as well as the loss between the reflected light component and the illumination light component. The objective function for underwater image enhancement is represented by a loss function, and the training of a network model for low-light underwater images is completed. ； ； ； ； ； in, This represents the underwater image predicted by the network. This represents the underwater images predicted by the network. Corresponding real underwater images, Indicates the mean squared error loss. It is the structural similarity index loss. This represents the reflected light component of the underwater image predicted by the network. This represents the underwater image illumination light component predicted by the network. This represents the reflected light component of a real underwater image. This represents the illumination light component of a real underwater image. This represents the absolute error loss.

7. The low-light underwater image enhancement method according to claim 1, characterized in that, The evaluation metrics in step 4 include peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). ； ； in, and These represent high-quality underwater images belonging to the model output. pixels and Corresponding real underwater images The pixels; This represents the maximum value of the color at a point in the image, with a value of 255. Indicates the number of samples in the training set. Indicates the current sample; ； ； ； in, yes average value, Is with Corresponding real underwater images The average value, yes variance yes The variance; L is the dynamic range of pixel values. and All represent preset hyperparameters, and their values ​​here are: =0.01, =0.03, and All are smoothing parameters.

8. A low-light underwater image enhancement model, characterized in that, Includes the following modules: Model building module: used to build low-light underwater image datasets and low-light underwater image enhancement network models; The low-light underwater image enhancement network model includes: a deep Retinex decomposition module, multiple cascaded feature encoders and feature decoders, and a feature fusion module; The process of low-light underwater image enhancement network model performing low-light underwater image enhancement is as follows: First, input a low-resolution, low-light image. The light source is decomposed based on the depth Retinex decomposition module to obtain the decomposed reflected light components. and illumination light components : For input low-light underwater images The features are concatenated to increase their dimensionality, resulting in initial decomposed features. Then, 9x9 convolution operations are used for feature enhancement, followed by five sets of 3x3 convolutions and ReLU activation functions for further feature enhancement and extraction. Finally, the obtained features are segmented to form reflected light components. and illumination light components : ； in, It is an input low-light underwater image. Connect the two vectors along the channel dimension. and It's a convolution operation. It is an activation function. It involves segmenting features. It is the reflection component. It is the component of illumination light; For the illumination light component After feature extraction, the input image is fused. Obtain the initial feature map For the initial feature map After processing by multiple cascaded feature encoders and feature decoders, the reflected light components are fused. The enhanced depth feature map is obtained. ; Finally, the feature fusion module is used to process the deep feature map. By fusing the images, high-quality underwater images can be obtained. ; Training module: Used to train the constructed low-light underwater image enhancement network model based on the low-light underwater image dataset; Prediction and Evaluation Module: This module uses a trained low-light underwater image enhancement network model to predict test samples of low-light underwater images, obtain high-quality underwater images, and evaluate them using evaluation metrics.

Images