Low-light underwater image enhancement method and system

Abstract

The invention discloses a low-light underwater image enhancement method and system, and the method comprises the steps: constructing a low-light underwater image data set, constructing a low-light underwater image enhancement network model, and carrying out the training of the constructed low-light underwater image enhancement network model based on the low-light underwater image data set, and finally, predicting a low-light underwater image test sample by using the trained low-light underwater image enhancement network model to obtain a high-quality underwater image, and evaluating by using an evaluation index. According to the scheme of the invention, specific scene limitation brought by a pure physical algorithm can be effectively relieved by adopting the Retienx decomposition module based on deep learning, and information interaction between coding layer features and decoding layer features can be promoted on a global level by utilizing the decoder feature module and the global illumination feature module. Therefore, the repairing capability of the model is greatly improved, the effectiveness of the method is proved through wide experiments, and the conclusion that the method provided by the invention has better performance can be obtained by carrying out experiments on a real low-light underwater image data set.

Description

Technical Field

[0001] The present invention belongs to the field of image enhancement, and in particular relates to a low-light underwater image enhancement method and system. Background Art

[0002] Under the condition of increasing pressure on land space and resources, marine resources are gradually becoming a new world center. Underwater images are important perceptual information for observing underwater environments. However, since the propagation attenuation of light in water is an uneven attenuation that depends on the wavelength characteristics, underwater images have serious color deviations. At the same time, particles in the water body reflect light, which makes the details of the imaging blurred and affects the quality of the image. These complex underwater environments lead to problems such as color deviation and high noise in the image. With the vigorous development of underwater detection technology, the depth that can be detected is getting larger and larger, but the visibility is low during deep-sea operations, resulting in not only color deviation in the underwater images taken, but also problems such as low image brightness and severe noise. At present, underwater image enhancement algorithms based on deep learning are difficult to effectively improve the visual effects of images in the face of complex degradation in low-light underwater images. Although low-light image enhancement algorithms in natural scenes can effectively correct the illumination of images and reduce noise interference in images, it is difficult to effectively solve problems such as color distortion caused by underwater imaging. In summary, although the current underwater image enhancement technology and low-illumination image enhancement algorithm can solve some problems of low-light underwater images, they cannot meet the effect requirements of low-light underwater image enhancement. Summary of the invention

[0003] In view of the above problems, the object of the present invention is to provide a low-light underwater image enhancement method and system, which pays attention to the color correction degree of the underwater image while paying attention to the illumination restoration of the low-light underwater image.

[0004] The specific technical solution for achieving the purpose of the present invention is:

[0005] A low-light underwater image enhancement method comprises the following steps:

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

[0007] Step 2: Build a low-light underwater image enhancement network model;

[0008] Step 3: Training 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 low-light underwater image test samples to obtain high-quality underwater images, and evaluate them using evaluation indicators.

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

[0011] Model building module: used to build a low-light underwater image dataset and a low-light underwater image enhancement network model;

[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: used to use the trained low-light underwater image enhancement network model to predict low-light underwater image test samples, obtain high-quality underwater images, and evaluate them using evaluation indicators.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

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

[0016] Specifically, we first apply a deep Retinex decomposition module to the input image to obtain the reflected light component and the illumination light component. The deep learning-based Retinex decomposition module can effectively alleviate the specific scene limitations brought by the pure physical algorithm.

[0017] Afterwards, the present invention performs a convolution operation on the illumination light component and fuses the original low-light underwater image input by splicing to obtain an initial feature map, where the lost details are supplemented for the features by splicing and fusion operations; then, the present invention applies multiple feature encoding modules and feature decoding modules to the initial features to fuse the reflected light components to obtain a depth feature map, the encoding feature module and the decoder feature module can efficiently capture the long-distance dependency between features by using the illumination light attention, and use the multi-scale fusion module to efficiently fuse the feature context information at different scales, and in addition, the decoder feature module can promote the information interaction between the encoding layer features and the decoding layer features at the global level by using the global illumination feature module, thereby greatly improving the repair ability of the model;

[0018] The present invention trains the network by minimizing the mean square 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 the reflected light component and the illumination light component, to improve the network's restoration ability. The effectiveness of the method is demonstrated through extensive experiments, and experiments on real low-light underwater image datasets show that the method proposed by the present invention has good performance.

[0019] The present invention is further described below in conjunction with specific implementation modes. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0021] Figure 2 4 is a low-light underwater image enhancement network framework diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] Example

[0023] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an kind" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0025] Unless otherwise specifically stated, the relative arrangement, numerical expressions and numerical values ​​of the parts and steps set forth in these embodiments do not limit the scope of the present application. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to the actual proportional relationship. The technology, method and equipment known to those of ordinary skill in the relevant field may not be discussed in detail, but in appropriate cases, the technology, method and equipment should be considered as a part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so that once a certain item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.

[0026] Combination Figure 1 , a low-light underwater image enhancement method, comprising the following steps:

[0027] Step 1: Construct a low-light underwater image dataset S = {(x 1 ,y 1 ),(x 2 ,y 2 ),...,(x L ,yL )};

[0028] Among them, x i represents a low-light underwater image, y i Represents x i The corresponding high-quality real underwater images, i = 1,…,L, L represents the number of samples in the dataset;

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

[0030] Step 2: Combine Figure 2 , build a low-light underwater image enhancement network model, including a deep Retinex decomposition module, multiple serially connected feature encoder and feature decoder modules, and a feature fusion module;

[0031] In this embodiment, the feature encoders include three and the feature decoder modules include two;

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

[0033] First, input a low-resolution dark image I l , based on the deep Retinex decomposition module, the light source is decomposed. The deep Retinex decomposition module includes splicing and dimensionality increase operations and multiple convolution operations. First, the input low-light underwater image I l The initial decomposition features are obtained by splicing to increase the dimension, and then the features are enhanced by 9*9 convolution operation, and then the features are enhanced and extracted by 5 groups of 3*3 convolution and Relu activation function. Finally, the decomposed reflected light component R and illumination light component L are obtained, which are as follows:

[0034] For the input low-light underwater image I l The initial decomposition features are obtained by splicing to increase the dimension, and then the features are enhanced by 9*9 convolution operation, and then by 5 groups of 3*3 convolution and Relu activation function. Finally, the obtained features are segmented to form the reflected light component R and the illumination light component L:

[0035]

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

[0037] After extracting the features of the illumination light component L, the input image I is fused l The initial feature map F is obtained. After being processed by three series-connected feature encoders and two feature decoders, the reflected light component R is fused to obtain the enhanced depth feature map F. d :

[0038] Perform 3*3 convolution feature extraction on the illumination light component L and fuse the input original underwater image features to generate the initial feature map F:

[0039] F=Conv 3*3 (L)⊙I l +I l

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

[0041] The plurality of serially connected feature encoders each include a 4*4 convolution unit and an illumination light attention unit, wherein the illumination light attention unit includes an illumination light attention layer and a multi-scale feature fusion layer to enhance the input features;

[0042] The processing of each feature encoder is:

[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 is the input feature map. The input feature of the initial feature coding layer is the illumination light component L. The input feature of each subsequent feature coding layer is the output feature of the upper feature coding layer. Conv 4*4 is the convolution operation, F L is the input feature of the illumination light attention unit, F R is the input reflected light feature, which is obtained by dimensionality reduction processing of the reflected light component R, LN is the layer normalization operation, and F c is the intermediate feature, F eis the output feature, 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 enhance the features output by the previous encoder feature module in turn and finally obtain the encoded feature output.

[0048] Among them, the RA illumination light attention layer is mainly used to enhance the long-distance dependencies between features. First, the input features are F LN After the row dimension is reshaped, a 1D convolution operation is used to aggregate the pixel-level cross-channel context to obtain the query F Q , key F k Sum value F V , for the input reflected light feature F R We also perform reshaping and 1D convolution operations, multiply the feature vector and value to get the latest value, and finally apply the self-attention rule to calculate the output feature F. a , the processing process is:

[0049]

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

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

[0052] The MFFB multi-scale feature fusion layer is mainly used to interactively fuse the enhanced features. First, the input feature F CLPerform convolution operation, then apply 3*3 and 5*5 convolution size deep convolution and Relu activation function to the features respectively, then concatenate these features, and apply 3*3 and 5*5 convolution size deep convolution and Relu activation function to the features again, and finally concatenate and apply convolution operation to obtain the output feature F M The specific processing process is:

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

[0054]

[0055] F M =Conv 1*1 ([F′ x ,F′ y ])

[0056] Among them, F CL Yes F C After layer normalization, the input features of Conv 1*1 is a convolution operation, DW 3*3 and DW 5*5 is a deep convolution operation, Chunk is a feature segmentation operation, [·] connects two vectors in the channel dimension, F M is the output feature;

[0057] The two serial feature decoders each include a 2*2 depth convolution layer, a feature fusion layer, a 1-dimensional convolution layer, a global illumination feature layer, and an illumination light attention layer. First, for the input feature F e Perform 2*2 depth convolution operation, and the fusion module concatenates the feature F′ of the encoding layer e The fusion feature is then fused, and then a 1-dimensional convolution operation and a global illumination feature module operation are performed on the fused feature. Finally, the illumination light attention module is used to perform attention on the feature and the reflected light feature F R Perform attention enhancement and finally obtain the output feature F d The remaining decoders sequentially enhance the features output by the feature module of the previous decoder to finally obtain the decoder feature output.

[0058] The processing of each feature decoder is:

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

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

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

[0062] Among them, F e It is the input feature. The input feature of the initial feature decoding layer is the output feature of the last feature encoding layer, and the input feature of the subsequent feature decoding layer is the output feature of the previous feature decoding layer. 2*2 is the depthwise convolution operation, F′ e is the output feature of the feature encoding layer at the symmetric level, [·] means connecting the two vectors in the channel dimension, Conv 1*1 is a convolution operation, FA represents the global illumination feature layer processing process, and F L is the input feature of the illumination light attention module, F R is the input reflected light feature, which is obtained by dimensionality reduction processing 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 is the intermediate feature, F d is the output feature;

[0063] Among them, in order to enhance the intermediate features of the fused coding layer features, the global illumination feature layer first performs the input feature F cn Perform adaptive average pooling convolution of different dimensions, perform 1D depth convolution on the features of H and W dimensions respectively, and then perform cross multiplication. Apply 3*3 convolution to the result of the cross multiplication to obtain the intermediate features. Perform 1D convolution on the features of C dimension and then perform Sigmoid activation. Then perform dot multiplication with the intermediate features, apply 1D convolution and Sigmoid activation function, and perform dot multiplication with the initial input features to obtain the final output feature F. f , the processing process is:

[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 (Fcn )))

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

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

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

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

[0070] Finally, based on the feature fusion module, the deep feature map F d The feature fusion module includes a 3*3 convolution and residual operation. First, a 3*3 convolution is used to fused the deep feature map F d Perform dimensionality reduction and then fuse the initial feature map F 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 is the acquired deep feature map, Conv 3*3 is the convolution operation, F is the initial feature map, I h It is a high quality clear underwater image enhancement image.

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

[0074] During the training process, the training objectives are 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;

[0075] Use the loss function to represent the objective function of underwater image enhancement and complete the training of the low-light underwater image network model

[0076]

[0077] L r =||I h -I gt || 2

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

[0079]

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

[0081] Step 4: Use the trained low-light underwater image enhancement network model to predict the low-light underwater image test samples to obtain high-quality underwater images, and use evaluation indicators to evaluate them;

[0082] The evaluation indicators include peak signal-to-noise ratio PSNR and structural similarity SSIM:

[0083]

[0084]

[0085] Among them, I i h and I i gt They represent the high-quality underwater images I belonging to the model output. h The pixel sum and I h The corresponding real underwater image I gt Pixels; MAX IIndicates the maximum value of the color of the image point, which is generally 255. n represents the number of training set samples, and i represents the current sample.

[0086] c 1 =(k 1 L) 2

[0087] c 2 =(k 2 L) 2

[0088]

[0089] Among them, μ Ih is the high-quality underwater image I output by the model h The average value of Yes and I h The corresponding real underwater image I gt The average value of isI h The variance of isI gt variance; L is the dynamic range of pixel values, k 1 and k 2 All represent preset hyperparameters, and the values ​​here are: k 1 =0.01, k 2 =0.03, Represents Y pred Standard Deviation and Y gt Standard Deviation The product of 1 and c 2 Both are smoothing parameters. Structural similarity SSIM compares the brightness, contrast and structural characteristics between images.

[0090] The test data is enhanced to obtain high-quality underwater images, which are evaluated using the customized metrics mentioned above. The effectiveness of this method is demonstrated through extensive experiments and applied to low-light underwater images.

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

[0092] Table 1 Results of low-light underwater image enhancement indicators

[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 a low-light underwater image dataset and a low-light underwater image enhancement network model;

[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: used to use the trained low-light underwater image enhancement network model to predict low-light underwater image test samples, obtain high-quality underwater images, and evaluate them using evaluation indicators.

[0098] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A low-light underwater image enhancement method, characterized in that: The following steps are involved: Step 1: Build a low-light underwater image dataset; Step 2: Build a low-light underwater image enhancement network model; Step 3: Training 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 low-light underwater image test samples to obtain high-quality underwater images, and evaluate them using evaluation indicators.

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 S = {(x1, y1), (x2, y2), ..., (x L ,y L )}; Among them, x i represents a low-light underwater image, y i Represents x i The corresponding high-quality real underwater images, i = 1, ..., L, L represents the number of samples in the dataset; The maximum and minimum normalization method is used to control the data range in the low-light underwater image dataset between 0 and 1.

3. The low-light underwater image enhancement method according to claim 1, characterized in that: The low-light underwater image enhancement network model in step 1 includes: a deep Retinex decomposition module, a plurality of serially connected feature encoders and feature decoder modules, and a feature fusion module; The process of low-light underwater image enhancement by the low-light underwater image enhancement network model is as follows: First, input a low-resolution dark image I l , perform light source decomposition based on the deep Retinex decomposition module to obtain the decomposed reflected light component R and illumination light component L; After extracting the features of the illumination light component L, the input image I is fused l The initial feature map F is obtained. After being processed by multiple series-connected feature encoders and feature decoders, the reflected light component R is fused to obtain the enhanced depth feature map F. d ; Finally, based on the feature fusion module, the deep feature map F d Fusion is performed to obtain high-quality underwater image I h .

4. The low-light underwater image enhancement method according to claim 3, characterized in that: The process of light source decomposition based on the deep Retinex decomposition module is as follows: For the input low-light underwater image I l The initial decomposition features are obtained by splicing to increase the dimension, and then the features are enhanced by 9*9 convolution operation, and then by 5 groups of 3*3 convolution and Relu activation function. Finally, the obtained features are segmented to form the reflected light component R and the illumination light component L: Among them, I l is the input low-light underwater image, [·] concatenates the two vectors in the channel dimension, Conv 9*9 and Conv 3*3 It is a convolution operation, Relu is an activation function, Chunk is a feature cutting operation, R is the reflection component, and L is the illumination component.

5. The low-light underwater image enhancement method according to claim 3, characterized in that: The enhanced depth feature map F d The process is: Perform 3*3 convolution feature extraction on the illumination light component L and fuse the input original underwater image features to generate the initial feature map F: F=Conv 3*3 (L)⊙I l +I l Where L is the illumination light component, I l is the input low-light underwater image, Conv 3*3 is the convolution operation and F is the initial feature map; The plurality of serially connected feature encoders each include a 4*4 convolution unit and an illumination light attention unit, wherein the illumination light attention unit includes an illumination light attention layer and a multi-scale feature fusion layer to enhance the input features; The processing of each feature encoder is: F L =Conv 4*4 (F t ), F c =RA(LN(F L ), F R )+F L F e =MFFB(LN(F c ))+F c Among them, F t is the input feature map. The input feature of the initial feature coding layer is the illumination light component L. The input feature of each subsequent feature coding layer is the output feature of the upper feature coding layer. Conv 4*4 is the convolution operation, F L is the input feature of the illumination light attention unit, F R is the input reflected light feature, which is obtained by dimensionality reduction processing of the reflected light component R, LN is the layer normalization operation, and F c is the intermediate feature, F e is the output feature, RA represents the illumination light attention layer processing process, and MFFB represents the multi-scale feature fusion layer processing process; The multiple serially connected feature decoders each include a 2*2 depth convolution layer, a feature fusion layer, a 1-dimensional convolution layer, a global illumination feature layer and an illumination light attention layer; The processing of each feature decoder is: F L =FA(Conv 1*1 ([DW 2*2 (F e ),F′ e ]) F C =RA(LN(F L ), F R )+F L F d =MFFB(LN(F C ))+F c Among them, F e It is the input feature. The input feature of the initial feature decoding layer is the output feature of the last feature encoding layer, and the input feature of the subsequent feature decoding layer is the output feature of the previous feature decoding layer. 2*2 is the depthwise convolution operation, F′ e is the output feature of the feature encoding layer at the symmetric level, [·] means connecting the two vectors in the channel dimension, Conv 1*1 is a convolution operation, FA represents the global illumination feature layer processing process, and F L is the input feature of the illumination light attention module, F R is the input reflected light feature, which is obtained by dimensionality reduction processing 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 is the intermediate feature, F d is the output feature.

6. The low-light underwater image enhancement method according to claim 5, characterized in that: The illumination light attention layer processing process in the feature encoder is: F a =Conv 1*1 (F V ⊙F R ·Softmax(F Q ·F K ))+F LN Among them, F LN Yes F L The input feature after layer normalization operation, F′ LN Yes F LN The feature after reshaping the row dimension, F′ R Yes F R The input reflected light features after row dimension reshaping, Conv 1*1 is the convolution operation, F Q is a query, F K Yes key, F V is the value, F R is the reflected light feature F R The value of F LN is the input feature, F a is the output feature; The multi-scale feature fusion layer processing process in the feature encoder is: F n =Conv 1*1 (F CL ) F M =Conv 1*1 ([F′ x ,F′ y ]) Among them, F CL Yes F C After layer normalization, the input features of Conv 1*1 is a convolution operation, DW 3*3 and DW 5*5 is a deep convolution operation, Chunk is a feature segmentation operation, [·] connects two vectors in the channel dimension, F M is the output feature; The global illumination feature layer processing process in the feature decoder is: F H =DW 1*1 (AAP H (F cn )),F WM =DW 1*1 (AAP W (F cn )),F C =Sig(Conv 1*1 (AAP C (F cn , F f =Sig(CNN 1*1 (Conv 3*3 (F H ·F W )⊙F C ))⊙F cn Among them, F cn is the input feature map, AAP is the adaptive average pooling convolution, DW 1*1 is a deep convolution operation, Conv 1*1 and Conv 3*3 is the convolution operation, Sig is the Sigmoid activation function, and F f is the output feature; The illumination light attention processing process in the feature decoder is the same as the illumination light attention layer processing process in the feature encoder; The multi-scale feature fusion layer processing process in the feature decoder is the same as the multi-scale feature fusion layer processing process in the feature encoder.

7. The low-light underwater image enhancement method according to claim 3, characterized in that: The underwater image I h The acquisition process is: I h =Conv 3*3 (F d )+F Among them, F d is the acquired deep feature map, Conv 3*3 is the convolution operation, F is the initial feature map, I h It is a high quality clear underwater image enhancement image.

8. The low-light underwater image enhancement method according to claim 3, characterized in that: In the step 3, during the training process of the low-light underwater image enhancement network model, the training objectives are 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; Use the loss function to represent the objective function of underwater image enhancement and complete the training of the low-light underwater image network model L r =||I h -I gt ||2 L ssim =1-SSIM(I h ,I gt ) Among them, I h represents the underwater image predicted by the network, I gt Represents the underwater image I predicted by the network h The corresponding real underwater image, ||·||2 represents the mean square error loss, SSIM is the structural similarity index loss, R h represents the reflected light component of the underwater image predicted by the network, L h represents the illumination component of the underwater image predicted by the network, R gt Represents the reflected light component of the real underwater image, L gt represents the illumination light component of the real underwater image, and ||·||1 represents the absolute error loss.

9. The low-light underwater image enhancement method according to claim 3, characterized in that: The evaluation indicators in step 4 include peak signal-to-noise ratio PSNR and structural similarity SSIM: Among them, I i h and I i gt They represent the high-quality underwater images I belonging to the model output. h The pixel sum and I h The corresponding real underwater image I gt Pixels; MAX I Indicates the maximum value of the color of the image point, which is generally 255. n represents the number of training set samples, and i represents the current sample. c1=(k1L) 2 c2=(k2L) 2 in, is the high-quality underwater image I output by the model h The average value of Yes and I h The corresponding real underwater image I gt The average value of isI h The variance of isI gt The variance of L is the dynamic range of pixel values. k1 and k2 are both preset hyperparameters, and the values ​​here are: k1 = 0.01, k2 = 0.03, Represents Y pred Standard Deviation and Y gt Standard Deviation The product of , c1 and c2 are both smoothing parameters.

10. A low-light underwater image enhancement model, characterized in that: Includes the following modules: Model building module: used to build a low-light underwater image dataset and a low-light underwater image enhancement network model; 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: used to use the trained low-light underwater image enhancement network model to predict low-light underwater image test samples, obtain high-quality underwater images, and evaluate them using evaluation indicators.

Images