Quality evaluation method for underwater enhanced image

Abstract

The invention belongs to the technical field of image processing, and discloses a quality evaluation method for an underwater enhanced image, which comprises the following steps: S1, calculating a local normalized graph for an input enhanced underwater image, and measuring the naturalness of the enhanced underwater image; s2, calculating a gradient intensity graph of the input enhanced underwater image, and measuring the ambiguity of the enhanced underwater image; s3, performing convolution operation on the input enhanced underwater image to obtain a feature map, inputting the feature map into a channel space attention module to obtain a purified feature map, and measuring the color cast degree and the local distortion degree of the enhanced underwater image; s4, splicing the feature maps extracted in the steps S1 to S3, inputting the spliced feature maps into a channel attention module, and obtaining a quality score of the image through a multi-layer perceptron; and S5, predicting the quality of an image to be evaluated by using the trained model. The method does not need to refer to an original image, can obtain high prediction performance, is high in operation speed, and meets the application requirements of underwater operation.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing and relates to a quality evaluation method for underwater enhanced images. Background Art

[0002] Underwater imaging is widely used in underwater operations and is an important part of underwater operations. Due to the refraction and scattering of light by water, the quality of underwater imaging is low, often showing problems such as blue-green bias, blur, and low contrast, which cannot meet the needs of underwater operations. Therefore, the collected underwater images often need to be appropriately enhanced to improve their quality, so as to provide accurate information for underwater operations. However, due to the lack of supervision of accurate quality measurement methods, the existing underwater image enhancement methods are still relatively blind and cannot effectively improve the quality of underwater images.

[0003] In terms of natural image quality evaluation, some researchers have proposed the structural similarity method (SSIM), and some have used the change of visual saliency features to evaluate the quality of images, and some have used the gradient intensity features of images to evaluate the quality of images. Some researchers have proposed statistical modeling of the structure, naturalness and perception of images, and constructed an unsupervised quality evaluation method. In terms of underwater image quality evaluation, some researchers have proposed an underwater image quality evaluation index, which measures the color, contrast and saturation of the image respectively, and then evaluates the quality of the underwater image by linear weighted sum of the three. Some researchers have proposed an underwater image quality measurement method, which calculates the color, blur and contrast of the image respectively, and then evaluates the quality of the underwater image by linear weighting. Some researchers have proposed to build an underwater image quality evaluation model, such as an underwater image quality evaluation model based on channel spatial attention and Transformer, and some models measure the naturalness, sharpness and structure of enhanced underwater images.

[0004] Based on this, the present invention proposes a quality evaluation method for underwater enhanced images, which effectively guides the enhancement process of underwater images, thereby improving the efficiency of underwater operations. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a quality evaluation method for underwater enhanced images. The naturalness, blurriness, color deviation and local distortion of underwater enhanced images are characterized based on a deep neural network, and an image quality evaluation model is trained using an end-to-end training method to achieve accurate judgment of the quality of underwater enhanced images.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A quality evaluation method for underwater enhanced images comprises the following steps:

[0008] S1. Calculate the local normalized map:

[0009] The local normalized map is calculated for the input enhanced underwater image, and then the local normalized map is input into N cascaded residual blocks for feature extraction to obtain the feature map F N , measures the naturalness of underwater enhanced images;

[0010] S2. Calculate the gradient intensity map:

[0011] The gradient intensity map is calculated for the input enhanced underwater image, and then the gradient intensity map is input into N cascaded residual blocks for feature extraction to obtain the feature map F B , measures the blurriness of underwater enhanced images;

[0012] S3. Calculate feature map:

[0013] The input enhanced underwater image is convolved, and the feature map is input into the channel space attention module to obtain the purified feature map F', which measures the color deviation and local distortion of the underwater enhanced image;

[0014] S4. Feature stitching:

[0015] The feature maps extracted from S1-S3 are spliced, and the naturalness feature, blur feature, color cast feature and local distortion feature of the underwater enhanced image are integrated, input into the channel attention module, and passed through the multi-layer perceptron to obtain the image quality score;

[0016] S5. Image quality prediction:

[0017] Through S1-S4, a quality evaluation model for underwater enhanced images is constructed. The model contains three branches, namely, naturalness measurement branch, blur measurement branch, and color deviation and local distortion measurement branch. The features output by the three branches are concatenated, channel attention module and multi-layer perceptron to output the quality score of the image. End-to-end training is performed, and the trained model is used to predict the quality of the image to be evaluated.

[0018] Furthermore, the step S1 specifically includes: decomposing the input underwater enhanced image into three channels of red, green and blue, represented as R, G and B respectively, and then converting it into a grayscale image, represented as:

[0019]

[0020] in, Represents the grayscale image corresponding to the underwater image, and calculates Local normalized image of :

[0021]

[0022] in, represents the local normalized image, (i, j) represents the coordinate position of the pixel, μ(i, j) and σ(i, j) represent the local mean and variance, which are calculated by the following formula:

[0023]

[0024] Among them, s, t represent the offset of the coordinates, and S, T represent the maximum value of the offset;

[0025] Then, Input to the residual block for feature extraction to obtain the feature map F N , expressed as:

[0026]

[0027] Among them, ResBlock() represents N cascaded residual blocks.

[0028] Further, step S2, the specific method is as follows: using the Sobel operator to perform a convolution operation on the grayscale image to obtain horizontal and vertical gradient maps, and then calculating the gradient intensity map, and inputting the gradient intensity map into N cascaded residual blocks to obtain a feature map to measure the blur degree of the underwater enhanced image;

[0029] The step S2 specifically includes: using the horizontal and vertical Sobel operators to perform convolution operations on the image respectively:

[0030]

[0031] Among them, G x and G y Represent the horizontal and vertical gradient maps respectively, * represents the convolution operation, and then calculate the gradient intensity map:

[0032]

[0033] Among them, GM represents the gradient strength map; GM is input into N cascaded residual blocks to obtain the feature map F B :

[0034] F B =ResBlock(GM).

[0035] Furthermore, the specific method of step S3 is as follows: performing a convolution operation on the input underwater enhanced image to obtain a feature map, and then inputting the feature map into the channel space attention module, using the channel space attention mechanism to adjust and optimize the input feature map to obtain a purified feature map, and measuring the color deviation and local distortion of the underwater image;

[0036] Step S3 specifically includes: performing a convolution operation on the input underwater image using a 3×3 convolution kernel to obtain a feature map F:

[0037] F = conv 3×3 (I) Among them, conv 3×3 (·) represents the convolution operation, and the convolution kernel size is 3×3. Then F is input into the channel space attention module to purify the feature map, which is expressed as:

[0038]

[0039] Among them, the channel spatial attention module includes a channel attention module and a spatial attention module, F' represents the feature map adjusted by the channel attention module, and F" represents the feature map adjusted by the spatial attention module. Indicates element-wise multiplication operation, M c represents channel attention, M S represents spatial attention, M c (F) and M s (F) is calculated as follows:

[0040] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0041] M s (F) = σ(conv 7×7 [AvgPool(·); MaxPool(·)])

[0042] Among them, MLP represents multi-layer perceptron, AvgPool(·) and MaxPool(·) represent average pooling and maximum pooling, conv 7×7 (·) represents the convolution operation, the convolution kernel size is 7×7, and σ(·) represents the sigmoid function. Further, in step S4, the feature maps extracted from S1-S3 are concatenated together and then input into the channel attention module. Different feature maps are assigned different weights through the channel attention mechanism to optimize the feature maps. The obtained feature maps are input into the multi-layer perceptron to output the image quality score.

[0043] Specifically, the feature maps extracted from S1-S3 are concatenated to obtain the feature map F c :

[0044] F c =concat(F N ,F B ,F”)

[0045] Among them, concat(·) represents the feature map concatenation operation, which converts F cInput channel attention module to optimize feature map and obtain feature map F f :

[0046]

[0047] F f Input is a multi-layer perceptron, which outputs an image quality score:

[0048] q=MLP(F f )

[0049] Here, q represents the image quality score.

[0050] Furthermore, the quality evaluation model of the underwater enhanced image constructed by S1-S4 described in step S5 is end-to-end trained, and the quality of the image to be evaluated is predicted by using the trained model. The specific method is as follows: a number of underwater enhanced images with subjective quality marks are selected, and the quality scores of the images are predicted by using the image quality evaluation model constructed by S1-S4, and the model is end-to-end trained by using the mean square error. After the training is completed, the quality of the underwater enhanced image to be evaluated is predicted;

[0051] The mean square error is defined as:

[0052]

[0053] Among them, q i is the quality score of the i-th image predicted by the model, s i represents the subjective quality score of the i-th image, and the model is optimized using the Adam algorithm;

[0054] After the training is completed, for a pair of underwater enhanced images I to be evaluated, it is input into the trained model, and the model outputs its quality score, which is expressed as:

[0055] q=M(I)

[0056] Where M(·) represents the trained image quality assessment model, and q represents the quality score of I.

[0057] Compared with the prior art, the present invention has the following advantages:

[0058] The underwater enhanced image quality evaluation method described in the present invention realizes accurate judgment of the quality of underwater enhanced images and obtains better prediction performance than the prior art. The present invention calculates the local normalized image of the image to measure the naturalness of the underwater enhanced image, and uses residual learning to learn deep-level naturalness perception features, which measures image quality more accurately than traditional statistics-based methods (existing methods do not consider the naturalness measurement of images, and the present invention proposes naturalness measurement for the first time in the quality evaluation of underwater enhanced images). Similarly, the gradient intensity image is calculated, and the deep-level blur characteristics are learned using residual learning to more accurately measure the blur degree of the image.

[0059] In the quality evaluation of underwater enhanced images, this paper proposes for the first time to use the channel space attention mechanism to measure the degree of color deviation and local distortion of underwater enhanced images. By fusing the naturalness features, blur features, color deviation features and local distortion features of underwater enhanced images, accurate measurement of the quality of underwater enhanced images is achieved.

[0060] In addition, the present invention does not need to refer to the original underwater image, the method runs quickly, and meets the application requirements of underwater operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0062] Figure 1 It is a model structure diagram of the present invention;

[0063] Figure 2 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0064] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0065] The embodiments of the present invention respectively characterize the naturalness, blurriness, channel and local distortion degree of the underwater enhanced image, use a multi-layer perceptron to perform feature fusion and quality score prediction, and use an end-to-end training method to train the image quality evaluation model. After the training is completed, the model is used to make an accurate judgment on the quality of the underwater enhanced image.

[0066] like Figure 1 , Figure 2 As shown, an embodiment of the present invention provides a method for evaluating the quality of an underwater enhanced image, the method comprising:

[0067] S1. Calculate the local normalized map:

[0068] The local normalized map is calculated for the input enhanced underwater image, and then the local normalized map is input into N (for example, 5) cascaded residual blocks for feature extraction to obtain the feature map F N , measures the naturalness of underwater enhanced images.

[0069] S2. Calculate the gradient intensity map:

[0070] The gradient intensity map is calculated for the input enhanced underwater image, and then the gradient intensity map is input into N (for example, 5) cascaded residual blocks for feature extraction to obtain the feature map F B , measure the blur of underwater enhanced image; S3, calculate feature map:

[0071] The input enhanced underwater image is convolved, and the feature map is input into the channel space attention module to obtain the purified feature map F', which measures the color deviation and local distortion of the underwater enhanced image;

[0072] S4. Feature stitching:

[0073] The feature maps extracted from S1-S3 are spliced, and the naturalness feature, blur feature, color cast feature and local distortion feature of the underwater enhanced image are integrated, input into the channel attention module, and passed through the multi-layer perceptron to obtain the image quality score;

[0074] S5. Image quality prediction:

[0075] Through S1-S4, a quality evaluation model for underwater enhanced images is constructed. The model contains three branches, namely, naturalness measurement branch, blur measurement branch, and color deviation and local distortion measurement branch. The features output by the three branches are concatenated, channel attention module and multi-layer perceptron to output the quality score of the image. End-to-end training is performed, and the trained model is used to predict the quality of the image to be evaluated.

[0076] This method does not require reference to the original image and can achieve higher prediction performance while running fast, meeting the application requirements of underwater operations.

[0077] In one embodiment, the specific implementation process and details of the above-mentioned underwater enhanced image quality assessment method are as follows:

[0078] Specifically, S1 decomposes the input underwater enhanced image into three channels: red, green, and blue, represented as R, G, and B, respectively, and then converts it into a grayscale image, represented as:

[0079]

[0080] in, Represents the grayscale image corresponding to the underwater image, and calculates Local normalized image of :

[0081]

[0082] in, represents the local normalized image, (i, j) represents the coordinate position of the pixel, μ(i, j) and σ(i, j) represent the local mean and variance, which are calculated by the following formula:

[0083]

[0084] Among them, s, t represent the offset of the coordinates, S, T represent the maximum value of the offset, and S and T are set to 5.

[0085] Then, Input to the residual block for feature extraction to obtain the feature map F N , expressed as:

[0086]

[0087] Among them, F N Represents the feature map output by the residual block, and ResBlock() represents the residual block.

[0088] The present invention proposes a naturalness metric for the first time in the quality evaluation of underwater enhanced images. Existing methods do not consider the naturalness metric of images. How to accurately measure the naturalness of images is one of the problems solved by the present invention. The present invention proposes to achieve this by using residual learning.

[0089] Specifically, S2 uses the Sobel operator to perform convolution operations on the grayscale image to obtain horizontal and vertical gradient maps, and then calculates the gradient intensity map. The gradient intensity map is input into N cascaded residual blocks to obtain a feature map to measure the blur degree of the underwater enhanced image.

[0090] Specifically, the horizontal and vertical Sobel operators are used to perform convolution operations on the input underwater enhanced image respectively:

[0091] Among them, G x and G y Represent the horizontal and vertical gradient maps respectively, * represents the convolution operation, and then calculate the gradient intensity map:

[0092]

[0093] Among them, GM represents the gradient strength map. GM is input into 5 cascaded residual blocks to obtain the feature map F B :

[0094] F B =ResBlock(GM)

[0095] Among them, F B is the obtained feature map.

[0096] Specifically, S3 performs a convolution operation on the input underwater enhanced image to obtain a feature map, and then inputs the feature map into the channel space attention module. The channel space attention mechanism is used to adjust and optimize the input feature map to obtain a purified feature map to measure the color deviation and local distortion of the underwater image.

[0097] As an embodiment, a convolution operation is performed on the input underwater enhanced image using a 3×3 convolution kernel to obtain a feature map F:

[0098] F = conv 3×3 (I)

[0099] Among them, conv 3×3 (·) represents the convolution operation, and the convolution kernel size is 3×3. Then F is input into the channel space attention module to purify the feature map:

[0100]

[0101] Among them, the channel spatial attention module includes a channel attention module and a spatial attention module, F' represents the feature map adjusted by the channel attention module, and F" represents the feature map adjusted by the spatial attention module. Indicates element-wise multiplication operation, M c represents channel attention, M S represents spatial attention, M c (F) and M s (F) is calculated as follows:

[0102] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0103] M s (F) = σ(conv 7×7 [AvgPool(·); MaxPool(·)])

[0104] Among them, MLP represents multi-layer perceptron, AvgPool(·) and MaxPool(·) represent average pooling and maximum pooling, conv 7×7 (·) represents the convolution operation, the convolution kernel size is 7×7, and σ(·) represents the sigmoid function, which is defined as:

[0105]

[0106] Here, x is the input variable of the function.

[0107] It should be noted that, in the underwater enhanced image quality evaluation, the present invention uses the channel space attention mechanism for the first time to measure the degree of color deviation and local distortion of the image.

[0108] Specifically, step S4 concatenates the feature maps extracted by S1-S3, and then inputs them into the channel attention module. Different feature maps are assigned different weights through the channel attention mechanism to optimize the feature maps. The obtained feature maps are input into the multi-layer perceptron to output the image quality score.

[0109] This step achieves accurate measurement of underwater enhanced image quality by splicing feature maps and integrating the naturalness, blur, color shift, and local distortion features of underwater enhanced images. After feature splicing, the channel attention mechanism is introduced to assign different weights to different feature maps to improve the accuracy of image quality evaluation.

[0110] Specifically, the feature maps extracted in the above steps are concatenated to obtain the feature map F c :

[0111] F c =concat(F N ,F B ,F”)

[0112] Among them, concat(·) represents the feature map concatenation operation, which converts F c Input channel attention module to optimize feature map and obtain feature map F f :

[0113]

[0114] F f Input is a multi-layer perceptron, which outputs an image quality score:

[0115] q=MLP(F f )

[0116] Here, q represents the image quality score.

[0117] Step S5, performing end-to-end training on the underwater enhanced image quality evaluation model constructed in S1-S4, and using the trained model to predict the image quality to be evaluated.

[0118] Specifically, 1000 underwater enhanced images with subjective quality marks are selected, and the image quality evaluation model constructed in the above steps is used to predict the quality scores of these 1000 images. The model is trained end-to-end using the mean square error, and the mean square error is defined as:

[0119]

[0120] Among them, q i is the quality score of the i-th image predicted by the model, s i represents the subjective quality score of the i-th image. The model is optimized using the Adam algorithm, and the learning rate is set to 1e -4 , the number of samples used for each iteration training is set to 32.

[0121] After the training is completed, for a pair of underwater enhanced images I to be evaluated, it is input into the trained model, and the model outputs its quality score, which is expressed as:

[0122] q=M(I)

[0123] Where M(·) represents the trained image quality assessment model, and q represents the quality score of I.

[0124] In a specific embodiment of the present invention, the performance of the present invention is tested on the SOTA library (the paper “Generation-based joint luminance-chrominance learning for underwater image quality assessment” published by Wang et al. in IEEE Transactions on Circuits and Systems for Video Technology, Vol. 33, No. 3, pp. 1123 to 1139), and the performance of the algorithm is measured using three indicators: Spearman rank correlation coefficient (SRCC), Pearson linear correlation coefficient (PLCC) and root mean square error (RMSE). The experimental results are shown in Table 1, wherein the higher the SRCC and PLCC, and the lower the RMSE, the better the prediction performance.

[0125] Table 1 Performance comparison on SOTA database

[0126] Performance Indicators UCIQE UIQM CCF FDUM GQ The present invention SRCC 0.1425 0.1453 0.4337 0.0425 0.7996 0.8484 PLCC 0.4573 0.1044 0.4840 0.2495 0.8377 0.8766 RMSE 11.2807 12.6184 11.0895 12.2805 6.9278 6.1051

[0127] Table 1 compares the latest underwater image quality evaluation methods, including UCIQE, UIQM, CCF, FDUM, and GLCQE. It can be seen that the present invention achieves higher prediction performance than similar methods, which proves the effectiveness and superiority of the present invention.

[0128] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A quality assessment method for underwater enhanced images, characterized in that: The following steps are involved: S1. Calculate the local normalized map: The local normalized map is calculated for the input enhanced underwater image, and then the local normalized map is input into N cascaded residual blocks for feature extraction to obtain the feature map F N , measures the naturalness of underwater enhanced images; S2. Calculate the gradient intensity map: The gradient intensity map is calculated for the input enhanced underwater image, and then the gradient intensity map is input into N cascaded residual blocks for feature extraction to obtain the feature map F B , measures the blurriness of underwater enhanced images; S3. Calculate feature map: The input enhanced underwater image is convolved, and the feature map is input into the channel space attention module to obtain the purified feature map F', which measures the color deviation and local distortion of the underwater enhanced image; S4. Feature stitching: The feature maps extracted from S1-S3 are spliced, and the naturalness feature, blur feature, color cast feature and local distortion feature of the underwater enhanced image are integrated, input into the channel attention module, and passed through the multi-layer perceptron to obtain the image quality score; S5. Image quality prediction: Through S1-S4, a quality evaluation model for underwater enhanced images is constructed. The model contains three branches, namely, naturalness measurement branch, blur measurement branch, and color deviation and local distortion measurement branch. The features output by the three branches are concatenated, channel attention module and multi-layer perceptron to output the quality score of the image. End-to-end training is performed, and the trained model is used to predict the quality of the image to be evaluated.

2. The method for evaluating the quality of underwater enhanced images according to claim 1, characterized in that: The step S1 specifically includes: decomposing the input underwater enhanced image into three channels of red, green and blue, represented as R, G and B respectively, and then converting it into a grayscale image, represented as: in, Represents the grayscale image corresponding to the underwater image, and calculates Local normalized image of : in, represents the local normalized image, (i, j) represents the coordinate position of the pixel, μ(i, j) and σ(i, j) represent the local mean and variance, which are calculated by the following formula: Among them, s, t represent the offset of the coordinates, and S, T represent the maximum value of the offset; Then, Input to the residual block for feature extraction to obtain the feature map F N , expressed as: Among them, ResBlock() represents N cascaded residual blocks.

3. The method for evaluating the quality of underwater enhanced images according to claim 1, characterized in that: Step S2, the specific method is as follows: use the Sobel operator to perform convolution operation on the grayscale image to obtain horizontal and vertical gradient maps, then calculate the gradient intensity map, input the gradient intensity map into N cascaded residual blocks, obtain a feature map, and measure the blur degree of the underwater enhanced image; The step S2 specifically includes: using the horizontal and vertical Sobel operators to perform convolution operations on the image respectively: Among them, G x and G y Represent the horizontal and vertical gradient maps respectively, * represents the convolution operation, and then calculate the gradient intensity map: Among them, GM represents the gradient strength map; GM is input into N cascaded residual blocks to obtain the feature map F B : F B =ResBlock(GM)。 4. The method for evaluating the quality of underwater enhanced images according to claim 1, characterized in that: The specific method of step S3 is as follows: performing a convolution operation on the input underwater enhanced image to obtain a feature map, and then inputting the feature map into the channel space attention module, using the channel space attention mechanism to adjust and optimize the input feature map to obtain a purified feature map, and measure the color deviation and local distortion of the underwater image; Step S3 specifically includes: performing a convolution operation on the input underwater image using a 3×3 convolution kernel to obtain a feature map F: F=conv 3×3 (I) Among them, conv 3×3 (·) represents the convolution operation, and the convolution kernel size is 3×3. Then F is input into the channel space attention module to purify the feature map, which is expressed as: Among them, the channel spatial attention module includes a channel attention module and a spatial attention module, F' represents the feature map adjusted by the channel attention module, and F" represents the feature map adjusted by the spatial attention module. represents the element-wise multiplication operation, Mc represents the channel attention, MS represents the spatial attention, and M c (F) and M s (F) is calculated as follows: M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) M s (F)=σ(conv 7×7 [AvgPool(·);MaxPool(·)]) Among them, MLP represents multi-layer perceptron, AvgPool(·) and MaxPool(·) represent average pooling and maximum pooling, conv 7×7 (·) represents the convolution operation, the convolution kernel size is 7×7, and σ(·) represents the sigmoid function.

5. The method for evaluating the quality of underwater enhanced images according to claim 1, characterized in that: In step S4, the feature maps extracted by S1-S3 are concatenated together and then input into the channel attention module. Different feature maps are assigned different weights through the channel attention mechanism to optimize the feature maps. The obtained feature maps are input into the multi-layer perceptron to output the image quality score. Specifically, the feature maps extracted from S1-S3 are concatenated to obtain the feature map F c : F c =concat(F N ,F B ,F”) Among them, concat(·) represents the feature map concatenation operation, which converts F c Input channel attention module to optimize feature map and obtain feature map F f : F f Input is a multi-layer perceptron, which outputs an image quality score: q=MLP(F f ) Here, q represents the image quality score.

6. The method for evaluating the quality of underwater enhanced images according to claim 1, characterized in that: The quality evaluation model of the underwater enhanced image constructed by S1-S4 described in step S5 is end-to-end trained, and the quality of the image to be evaluated is predicted by using the trained model. The specific method is as follows: select a number of underwater enhanced images with subjective quality marks, use the image quality evaluation model constructed by S1-S4 to predict the quality scores of the several images, use the mean square error to train the model end-to-end, and after the training is completed, predict the quality of the underwater enhanced image to be evaluated; The mean square error is defined as: Among them, q i is the quality score of the i-th image predicted by the model, s i represents the subjective quality score of the i-th image, and the model is optimized using the Adam algorithm; After the training is completed, for a pair of underwater enhanced images I to be evaluated, it is input into the trained model, and the model outputs its quality score, which is expressed as: q=M(I) Wherein, M(·) represents the trained image quality assessment model, and q represents the quality score of image I.

Images