An underwater image quality evaluation method based on color space multi-feature fusion

Abstract

The application discloses a kind of underwater image quality evaluation methods based on color space multi-feature fusion, comprising: color space conversion is carried out to underwater image, and image is converted from RGB color space to CIELab color space;Image data feature extraction is carried out on the underwater image obtained after color space conversion on L component, AB component, respectively obtain L component feature, AB component feature, combine L component feature with AB component feature, form final feature vector;According to final feature vector, image quality evaluation model is established using machine learning method support regression vector, and the quality evaluation of underwater image is obtained by image quality evaluation model.The underwater image quality evaluation method based on image structure and morphological information and SVR is used in the application, the quality of underwater image can be accurately evaluated, has the advantages of objective, fast and strong stability, and provides scientific basis for the quality evaluation of underwater image.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of image processing, and relates to underwater image quality evaluation, in particular to an underwater image quality evaluation method based on color space multi-feature fusion. BACKGROUND

[0002] The underwater world contains rich resources, and underwater images play an important guiding role in the exploration and development of the underwater world, such as playing a key role in the fields of underwater archaeology, marine ecological research, underwater target detection and tracking. However, because of the influence of water absorption and scattering of light and noise and other factors, the underwater image quality has problems such as color deviation, low contrast and poor definition, which reduces the information amount of the underwater image and affects its accuracy in practical application. With the attention of scholars to the field of underwater vision research, the underwater image obtained through image processing technology has become a research hotspot and difficulty in recent years to improve the visual quality of the underwater image. However, the lack of a perfect underwater image quality evaluation system in this field will affect the selection of original underwater images and underwater image sharpening algorithms in practical applications, thereby limiting the development and progress of this research field. Therefore, designing an effective and robust underwater image quality evaluation method is a research hotspot and difficulty in the field of underwater image processing, and designing an effective underwater image quality evaluation method has important research significance and is currently urgently needed in this research field.

[0003] At present, underwater image quality evaluation is divided into subjective evaluation and objective evaluation. Subjective evaluation mainly evaluates image quality through human subjective perception, and this method depends on the experience and subjective feeling of professionals, is time-consuming and costly, and is not conducive to efficient and large-scale quality evaluation of underwater images. Objective evaluation uses computer simulation of human visual perception to evaluate image quality, and its advantage is simple operation and low cost. According to the amount of information needed, it can be divided into full-reference image quality method, partial-reference image quality evaluation method and no-reference image quality evaluation method. Full-reference image quality evaluation method needs to provide all reference information of the original image, and calculates the difference between the original image and the distorted image to obtain the quality of the distorted image. Partial-reference image quality evaluation method needs to provide part of the feature information of the original image as reference to evaluate the quality of the distorted image. Due to the particularity and complexity of the underwater environment, it is difficult to obtain a clear reference image underwater, so the no-reference image quality evaluation method is the best choice for underwater image quality objective evaluation. SUMMARY

[0004] Purpose of the invention: To overcome the shortcomings of existing technologies, this invention provides an underwater image quality assessment method based on multi-feature fusion in color space. Employing an underwater image quality assessment method based on image structure and morphological information and SVR, this method can accurately evaluate the quality of underwater images and has the advantages of objectivity, speed, and strong stability, providing a scientific basis for underwater image quality assessment.

[0005] Technical Solution: To achieve the above objectives, this invention provides an underwater image quality assessment method based on multi-feature fusion in color space, comprising the following steps:

[0006] S1: Convert the color space of the underwater image from RGB color space to CIELab color space;

[0007] S2: For the underwater image obtained after color space conversion, image data features are extracted in the L component and AB component, respectively. The L component features and AB component features are then combined to form the final feature vector.

[0008] S3: Based on the final feature vector, a machine learning method is used to establish an image quality assessment model using Support Vector Regression (SVR), and the quality assessment of the underwater image is obtained through the image quality assessment model.

[0009] Furthermore, in the CIELab color space of step S1, the L component is used to represent the brightness of the pixel, with a value range of [0, 100], representing pure black to pure white; a represents the range from red to green, with a value range of [127, -128]; b represents the range from yellow to blue, with a value range of [127, -128].

[0010] Furthermore, in step S1, the RGB color space cannot be directly converted to the CIELab color space and needs to be converted to the XYZ color space. The specific conversion process is as follows: first convert the RGB color space to the XYZ color space, and then convert the XYZ color space to the CIELab color space.

[0011] Furthermore, in step S1, the image is converted to the XYZ color space using the following formula:

[0012]

[0013] The formula for converting from XYZ color space to CIELab color space is as follows:

[0014] L * =116f(Y / Y) n )-16

[0015] a * =500[f(X / X) n )-f(Y / Y n )]

[0016] b * =200[f(Y / Y) n )-f(Z / Z n )]

[0017]

[0018] Among them, X n Y n Z n The CIE XYZ tri-color stimulus values ​​[0.9504, 1.0000, 1.0888] were used to simulate midday sunlight, with reference white point values.

[0019] Further, the L component features in step S2 include four features f1-f4. The specific extraction method is as follows: the L component image is downsampled, the MSCN coefficients are calculated between the L component image and its downsampled image to obtain the MSCN image, and the rotation-invariant LBP operator based on consistency is used for processing to obtain the LBP histogram as feature f1. The two MSCN images are fitted with an asymmetric generalized Gaussian distribution (AGGD) model and the kurtosis and skewness are calculated to obtain the morphological feature f2. The first to fourth order moment statistics are calculated on the L component image to obtain the moment statistics feature f3. The corresponding histogram is obtained according to the L component image. The range of the pixels in the image is [0, 100]. The frequency of these numbers is counted, and the group interval is 10. The resulting histogram is the brightness histogram feature f4.

[0020] Furthermore, the AB component features in step S2 include two features, f5 and f6, which are extracted using the following method:

[0021] The moment statistics feature f5 is obtained by calculating the first to fourth order moment statistics on the AB component plot;

[0022] Five color description maps were constructed based on the AB component map; they are:

[0023]

[0024] AGGD fitting was performed on five color description maps, and kurtosis and skewness were calculated. The morphological feature f6 was obtained. The specific process is as follows:

[0025] AGGD fitting was performed on the five color description maps to determine the shape parameter α and the left and right scale parameters. The mean η of the distribution;

[0026] The skewness S and kurtosis K are calculated from the five color description maps using the formula.

[0027] The shape parameter α and the left and right scale parameters The distribution's mean η, skewness S, and kurtosis K are combined as characteristic f6.

[0028] Furthermore, the extraction process of feature f1 in step S2 is as follows:

[0029] A1: Downsample the L component image. For image I (M×N), downsample it by a factor of 2 to obtain a downsampled image with a resolution of (M / 2)*(M / 2).

[0030] A2: Calculate the MSCN coefficients for the L component map and its downsampled map.

[0031] For an image I (M×N), after removing the mean and normalizing it, we obtain its MSCN (Mean Subtracted Contrast Normalized), i.e.

[0032]

[0033] Where: μ(x, y) is the result after Gaussian filtering, σ(x, y) is the standard deviation i∈{1, 2, ..., M}, j∈{1, 2, 3, ..., N}, M and N are the height and width of the image respectively, C is a constant 1 to prevent the denominator from being 0, and the variables μ(i, j) and σ(i, j) are defined as follows:

[0034]

[0035]

[0036] w={w k，l |k=-K,…,K, l=-L,…,L}

[0037] In the formula, w is a two-dimensional cyclic symmetric Gaussian weight function sampled from 3 standard deviations (K=L=3) and rescaled to a unit volume;

[0038] A3: Apply the consistency-based rotation-invariant local two-mode (LBP) operator to two MSCN graphs to calculate the relationship between each pixel. Let the value of a certain pixel (i, j) be m. ij Then the pixel values ​​of its P adjacent pixels within its radius r are m0, m1, ..., m p-1First, the consistency of pixel value relationships is statistically analyzed. If the pixel value of an adjacent pixel is greater than that of the center pixel, it is marked in red; otherwise, it is marked in white. If three consecutive pixels adjacent to the center pixel (i, j) have pixel values ​​greater than those of the center pixel (i, j), then the relationship value M_r of pixel (i, j) is determined. ij The value is 3; when there are non-contiguous pixels among the adjacent pixels of the center pixel (i, j) whose pixel values ​​are all greater than the center pixel value, the pixel relationship value M_r of pixel (i, j) is 3. ij It is p+1, which is 9; the pixel relationship value M_r ij The calculation method is shown in the following formula:

[0039]

[0040] In the formula, P is the number of neighboring pixels of pixel (i, j), and m p Let f be the pixel value of the p-th pixel (i, j); the calculation method of f in the formula is shown in the following equation; u ij Let u be the pixel consistency parameter between adjacent pixels of pixel (i, j), and its calculation method is shown in the following formula. From this formula, it can be seen that when u ij When u = 0, the pixel values ​​of all neighboring pixels of (i, j) are either less than or greater than the pixel value of (i, j); when u ij When u = 2, there are several consecutive neighboring pixels whose pixel values ​​are greater than the pixel value of (i, j); when u ij When the value is greater than 2, there are non-contiguous neighboring pixels whose pixel values ​​are greater than the pixel value of (i, j).

[0041]

[0042]

[0043] By calculating the relationship between each pixel in an image and its neighboring pixels, a relationship graph of the image can be obtained, which can effectively capture the structural information of the image.

[0044] Then, the histogram corresponding to the image is obtained through the relationship graph, and the relationship value M_r of the pixels in the image is obtained. ij Let f1 be an integer in the interval [1, 10]. By counting the probability of these integers, we can obtain the LBP histogram feature f1 of the underwater image.

[0045] Furthermore, the extraction process of feature f2 in step S2 is as follows:

[0046] B1: Perform AGGD fitting on the two obtained MSCN plots:

[0047] The AGGD model is shown in the following equation, where α controls the shape of the distribution. These are scale parameters, which control the distribution on both sides of the model, and η is the mean of the distribution;

[0048]

[0049] in:

[0050]

[0051]

[0052]

[0053]

[0054] B2: Calculate the skewness S and kurtosis K for the two MSCN plots. The specific calculation method is shown in the following formula:

[0055]

[0056]

[0057] B3: Shape parameter α and left and right scale parameters obtained by fitting the two MSCN images using AGGD. The mean η, skewness S, and kurtosis K of the distribution are used as morphological features f2.

[0058] The extraction process of moment statistics feature f3 in step S2 is as follows:

[0059] The first to fourth moment statistics of the L component are calculated using the following formulas: First moment (mean), second moment (variance), third moment (skewness), and fourth moment (kurtosis).

[0060]

[0061]

[0062]

[0063]

[0064] Where, p i，j Let f represent the i-th color component of the j-th pixel in the image, and N represent the number of pixels in the image; the four moment statistics obtained are used as features f3.

[0065] The extraction process of the brightness histogram feature f4 in step S2 is as follows:

[0066] C1: Obtain the histogram of the image from the L-component map; the value p of each pixel in the image. ij The integers are within the interval [0, 100].

[0067] C2: Calculate the probability of these integers, draw a histogram with a class interval of 10, normalize the histogram, and the result is the final brightness histogram feature f4.

[0068] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0069] 1. Compared with subjective image quality evaluation methods that are time-consuming, unstable, and dependent on human experience, this invention has the advantages of high efficiency and real-time performance.

[0070] 2. Compared with full-reference image quality assessment methods and half-reference image quality assessment methods, this invention does not require a reference image, which solves the problem of difficulty in obtaining reference images. Moreover, it has high consistency between subjective and objective evaluation results, is highly correlated with human visual perception, and has a wider range of application scenarios.

[0071] 3. Compared with existing no-reference image quality assessment methods that select only one feature, this invention considers multiple factors that affect underwater image quality, analyzes underwater image characteristics by component, mines multiple features related to underwater image quality, and extracts corresponding features on brightness and color components to quantify underwater image quality degradation, thus obtaining more accurate evaluation results.

[0072] 4. Compared with existing underwater image quality assessment methods that combine various metric components in a linear manner, this invention adopts a multi-feature fusion method and uses support vector regression to model the fused features, resulting in an underwater image quality assessment model with good performance and strong stability.

[0073] 5. This invention is geared towards underwater image quality assessment, and can provide a basis for judging underwater image quality and provide an effective and reliable reference for underwater image processing tasks, thus having high practical value. Attached Figure Description

[0074] Figure 1 : Flowchart of the method of the present invention;

[0075] Figure 2 The original diagram and component diagrams of this embodiment of the invention, wherein (a) is the original diagram, and (b), (c), and (d) are the L component diagram, A component diagram, and B component diagram, respectively;

[0076] Figure 3 : Pixel relationship value diagram of an embodiment of the present invention;

[0077] Figure 4 The MSCN diagram and pixel relationship diagram of this invention embodiment;

[0078] Figure 5 : LBP feature histogram of an embodiment of the present invention;

[0079] Figure 6 : Brightness feature histogram of an embodiment of the present invention;

[0080] Figure 7 The five color description diagrams constructed in the embodiments of the present invention;

[0081] Figure 8 Three reference figures from embodiments of the present invention;

[0082] Figure 9 The normal distribution fitting curve of the five color description maps constructed in this embodiment of the invention. Detailed Implementation

[0083] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0084] This invention provides a method for underwater image quality assessment based on multi-feature fusion in color space, such as... Figure 1 As shown, it includes the following steps:

[0085] S1: Convert the color space of the underwater image from RGB color space to CIELab color space;

[0086] S2: For the underwater image obtained after color space conversion, image data features are extracted in the L component and AB component, respectively. The L component features and AB component features are then combined to form the final feature vector.

[0087] S3: Based on the final feature vector, a machine learning method is used to establish an image quality assessment model using Support Vector Regression (SVR), and the quality assessment of the underwater image is obtained through the image quality assessment model.

[0088] In step S1:

[0089] While the RGB color space offers advantages in evaluating the quality of natural images, it performs poorly in underwater image quality assessment. The CIELab color space demonstrates a strong correlation between underwater visual quality and human perception. Therefore, color space conversion is necessary for underwater images, transforming them from RGB to CIELab. However, a direct conversion from RGB to CIELab requires the use of the XYZ color space. The specific conversion process involves first converting the RGB color space to the XYZ color space, and then converting the XYZ color space back to the CIELab color space.

[0090] In the CIELab color space, the L component is used to represent the brightness of a pixel, with a value range of [0, 100], representing pure black to pure white; a represents the range from red to green, with a value range of [127, -128]; b represents the range from yellow to blue, with a value range of [127, -128].

[0091] The image is converted to the XYZ color space using the following formula:

[0092]

[0093] The formula for converting from XYZ color space to CIELab color space is as follows:

[0094] L * =116f(Y / Y) n )-16

[0095] a * =500[f(X / X) n )-f(Y / Y n )]

[0096] b * =200[f(Y*Y n )-f(Z / Z n )]

[0097]

[0098] Among them, X n Y n Z n The CIE XYZ tri-color stimulus values ​​[0.9504, 1.0000, 1.0888] used as reference white point values ​​to simulate midday sunlight, with a correlated color temperature of 6504K. Figure 2 (b), (c), and (d) respectively show the image data of the L component, A component, and B component after color space conversion.

[0099] In step S2:

[0100] The L-component features include four features f1-f4. The specific extraction method is as follows: the L-component image is downsampled, and the MSCN coefficients are calculated between the L-component image and its downsampled image to obtain the MSCN image. The rotation-invariant LBP operator based on consistency is used for processing to obtain the LBP histogram as feature f1. The two MSCN images are fitted with an asymmetric generalized Gaussian distribution (AGGD) model, and the kurtosis and skewness are calculated to obtain the morphological feature f2. The first to fourth order moment statistics are calculated on the L-component image to obtain the moment statistics feature f3. The corresponding histogram is obtained based on the L-component image. The range of pixels in the image is [0, 100]. The frequency of these numbers is counted, and the group interval is 10. The resulting histogram is the brightness histogram feature f4.

[0101] The extraction process of feature f1 is as follows:

[0102] A1: Downsample the L component image. For image I (M×N), downsample it by a factor of 2 to obtain a downsampled image with a resolution of (M / 2)*(M / 2).

[0103] A2: Calculate the MSCN coefficients for the L component map and its downsampled map.

[0104] For an image I (M×N), after removing the mean and normalizing it, we obtain its MSCN (Mean Subtracted Contrast Normalized), i.e.

[0105]

[0106] Where: μ(x, y) is the result after Gaussian filtering, σ(x, y) is the standard deviation i∈{1, 2, ..., M}, j∈{1, 2, 3, ..., N}, M and N are the height and width of the image respectively, C is a constant 1 to prevent the denominator from being 0, and the variables μ(i, j) and σ(i, j) are defined as follows:

[0107]

[0108]

[0109] w={w k，l |k=-K,…,K, l=-L,…,L}

[0110] In the formula, w is a two-dimensional cyclic symmetric Gaussian weight function sampled from 3 standard deviations (K=L=3) and rescaled to a unit volume;

[0111] A3: Apply the consistency-based rotation-invariant local two-mode (LBP) operator to two MSCN graphs to calculate the relationship between each pixel. Let the value of a certain pixel (i, j) be m. ij Then the pixel values ​​of its P adjacent pixels within its radius r are m0, m1, ..., m p-1 ,like Figure 3 As shown in (a) in the figure. Figure 3 In the example, r = 1, P = 8. First, we check the consistency of the pixel value relationships. Figure 3 Examples (b) and (c) provide examples of consistent and inconsistent pixel values, respectively. If the pixel value of an adjacent pixel is greater than the pixel value of the center pixel, it is marked in red; otherwise, it is marked in white. Therefore, Figure 3 In (b), the pixel values ​​of three consecutive pixels adjacent to the center pixel (i, j) are all greater than the center pixel 56. The relationship value M_r between pixels (i, j) is... ij It is 3; Figure 3 In (c), if there are non-contiguous pixels among the adjacent pixels of the center pixel (i, j) whose pixel values ​​are all greater than the center pixel value, then the pixel relationship value M_r of pixel (i, j) is... ij It is p+1, which is 9; the pixel relationship value M_r ij The calculation method is shown in the following formula:

[0112]

[0113] In the formula, P is the number of neighboring pixels of pixel (i, j), and m p Let f be the pixel value of the p-th pixel (i, j); the calculation method of f in the formula is shown in the following equation; u ij Let u be the pixel consistency parameter between adjacent pixels of pixel (i, j), and its calculation method is shown in the following formula. From this formula, it can be seen that when u ij When u = 0, the pixel values ​​of all neighboring pixels of (i, j) are either less than or greater than the pixel value of (i, j); when u ij When u = 2, there are several consecutive neighboring pixels whose pixel values ​​are greater than the pixel value of (i, j); when u ij When the value is greater than 2, there are non-contiguous neighboring pixels whose pixel values ​​are greater than the pixel value of (i, j).

[0114]

[0115]

[0116]

[0117] By calculating the relationship between each pixel in an image and its neighboring pixels, a relationship graph of the image can be obtained, which can effectively capture the structural information of the image. Figure 4(a), (b), and (c) and (d) respectively show two MSCN diagrams and their relationship diagrams.

[0118] Then, the relationship diagram is used to derive... Figure 5 The histogram corresponding to the image shown shows the relationship value M_r of the pixels in the image. ij Let f1 be an integer in the interval [1, 10]. By counting the probability of these integers, we can obtain the LBP histogram feature f1 of the underwater image.

[0119] The extraction process of feature f2 is as follows:

[0120] B1: The two obtained MSCN diagrams ( Figure 4 AGGD fitting was performed on (a) and (b)).

[0121] The AGGD model is shown in the following equation, where α controls the shape of the distribution. These are scale parameters, which control the distribution on both sides of the model, and η is the mean of the distribution;

[0122]

[0123] in:

[0124]

[0125]

[0126]

[0127]

[0128] B2: Calculate the skewness S and kurtosis K for the two MSCN plots. The specific calculation method is shown in the following formula:

[0129]

[0130]

[0131] B3: Shape parameter α and left and right scale parameters obtained by fitting the two MSCN images using AGGD. The mean η, skewness S, and kurtosis K of the distribution are used as morphological features f2.

[0132] The extraction process of the moment statistic feature f3 is as follows:

[0133] The first to fourth moment statistics of the L component are calculated using the following formulas: First moment (mean), second moment (variance), third moment (skewness), and fourth moment (kurtosis).

[0134]

[0135]

[0136]

[0137]

[0138] Where, p i，j Let f represent the i-th color component of the j-th pixel in the image, and N represent the number of pixels in the image; the four moment statistics obtained are used as features f3.

[0139] The extraction process of brightness histogram feature f4 is as follows:

[0140] C1: Obtain the histogram of the image from the L-component map; the value p of each pixel in the image. ij The integers are within the interval [0, 100].

[0141] C2: Calculate the probability of these integers occurring, plot a histogram with a class interval of 10, and normalize the histogram. The result is as follows. Figure 6 As shown, this result is used as the final brightness histogram feature f4.

[0142] The AB component features include two features, f5 and f6, and the specific extraction method is as follows:

[0143] The moment statistics feature f5 is obtained by calculating the first to fourth order moment statistics on the AB component plot;

[0144] Five color description maps were constructed based on the AB component map; they are:

[0145]

[0146] in: It is the first-order partial derivative along the horizontal and vertical directions.

[0147] The five color description diagrams constructed are as follows Figure 7 (a), (b), (c), (d), and (e). To demonstrate the effectiveness of the constructed five color descriptors in scoring underwater image quality, the following were selected: Figure 8 The three reference figures (a), (b), and (c) show curves after fitting with a normal distribution function as follows: Figure 9 As shown, AGGD fitting is performed on five color description maps, and the kurtosis and skewness are calculated to obtain the morphological feature f6. The specific process is as follows:

[0148] AGGD fitting was performed on the five color description maps to determine the shape parameter α and the left and right scale parameters. The mean η of the distribution;

[0149] The skewness S and kurtosis K are calculated from the five color description maps using the formula.

[0150] The shape parameter α and the left and right scale parameters The distribution's mean η, skewness S, and kurtosis K are combined as characteristic f6.

[0151] In step S3:

[0152] A support regression vector model was used to establish an underwater image quality score evaluation model. Features f1-f6 were combined to obtain a feature vector, which was then trained using the LibSVM software package. The model learned the mapping relationship between the feature vector and the image quality score, thus achieving underwater image quality evaluation. The training set comprised 80% of the model during training. To avoid experimental errors caused by random sampling, the predicted train-test process was repeated 50 times.

[0153] Based on the above scheme, in order to verify the effectiveness and practical effect of the method of the present invention, this embodiment applies CMFUIQA (the method of the present invention) and other methods to the SAUD and LIVE datasets. The specific comparison results are shown in Tables 1 and 2. The optimal performance value corresponding to different evaluation indicators is highlighted in bold:

[0154] Table 1 - Performance comparison results on the SAUD database

[0155]

[0156] Table 2 - Performance comparison results on the UIED database

[0157]

[0158] To measure the general applicability of the method to different image types, the performance comparison results of CMFUIQA and other methods on classic databases are shown in Table 3:

[0159] Table 3 - Performance Comparison Results on Classic Databases

[0160]

[0161] Table 3 shows that the underwater image quality assessment algorithm has a certain degree of universality.

[0162] The effectiveness of the selected feature was evaluated by analyzing individual features. The results of ablation experiments on the CMFUIQA and SAUD databases are shown in Table 4.

[0163] Table 4 - Ablation experimental results from the SAUD database

[0164]

[0165]

[0166] It can be seen that the L component performs better than f1 to f4 individually, and the AB component fusion feature performs better than f5 and f6. The CMFUIQA method, which integrates L component and AB component features, is superior to the evaluation methods that use only L component features and only AB component features, and has higher subjective and objective consistency.

[0167] This demonstrates that the selection of these features is effective and can improve the image quality assessment. The experimental data above show that CMFUIQA can accurately evaluate the quality score of underwater images, exhibiting a high degree of consistency with subjective human evaluation results.

Claims

1. A method for evaluating the quality of underwater images based on multi-feature fusion in color space, characterized in that, Includes the following steps: S1: Convert the color space of the underwater image from RGB color space to CIELab color space; S2: For the underwater image obtained after color space conversion, image data features are extracted in the L component and AB component, respectively. The L component features and AB component features are then combined to form the final feature vector. S3: Based on the final feature vector, a machine learning method, support vector regression, is used to establish an underwater image quality assessment model, and the quality assessment of the underwater image is obtained through the image quality assessment model. In step S2, the L component features include four features f1-f4. The specific extraction method is as follows: the L component image is downsampled, and the MSCN coefficients are calculated between the L component image and its downsampled image to obtain the MSCN image. The rotation-invariant LBP operator based on consistency is used for processing to obtain the LBP histogram as feature f1. The two MSCN images are fitted with an asymmetric generalized Gaussian distribution model, and the kurtosis and skewness are calculated to obtain the morphological feature f2. The first to fourth order moment statistics are calculated on the L component image to obtain the moment statistics feature f3. The corresponding histogram is obtained based on the L component image. The range of the pixels in the image is [0, 100]. The frequency of these numbers is counted, and the group interval is 10. The resulting histogram is the brightness histogram feature f4. In step S2, the AB component features include two features, f5 and f6, which are extracted using the following method: The moment statistics feature f5 is obtained by calculating the first to fourth order moment statistics on the AB component plot; Five color description maps were constructed based on the AB component map; they are: AGGD fitting was performed on five color description maps and kurtosis and skewness were calculated to obtain morphological feature f6; The extraction process of feature f1 in step S2 is as follows: A1: Downsample the L component image. For image I (M×N), downsample it by a factor of 2 to obtain a downsampled image with a resolution of (M / 2)*(M / 2). A2: Calculate the MSCN coefficients for the L component map and its downsampled map. For an image I (M×N), after removing the mean and normalizing it, we obtain its MSCN, i.e. Where: μ(x, y) is the result after Gaussian filtering, σ(x, y) is the standard deviation i∈{1, 2, ..., M}, j∈{1, 2, 3, ..., N}, M and N are the height and width of the image respectively, C is a constant 1 to prevent the denominator from being 0, and the variables μ(i, j) and σ(i, j) are defined as follows: w={w k，l |k=-K,…,K,l=-L,…,L} In the formula, w is a two-dimensional cyclic symmetric Gaussian weight function sampled from 3 standard deviations (K=L=3) and rescaled to a unit volume; A3: Apply the consistency-based rotation-invariant local two-mode (LBP) operator to two MSCN graphs to calculate the relationship between each pixel. Let the value of a certain pixel (i, j) be m. ij Then the pixel values ​​of its P adjacent pixels within its radius r are m0, m1, ..., m p-1 First, the consistency of pixel value relationships is statistically analyzed. If the pixel value of an adjacent pixel is greater than that of the center pixel, it is marked in red; otherwise, it is marked in white. If three consecutive pixels adjacent to the center pixel (i, j) have pixel values ​​greater than those of the center pixel (i, j), then the relationship value M_r of pixel (i, j) is determined. ij The value is 3; when there are non-contiguous pixels among the adjacent pixels of the center pixel (i, j) whose pixel values ​​are all greater than the center pixel value, the pixel relationship value M_r of pixel (i, j) is 3. ij It is p+1, which is 9; the pixel relationship value M_r ij The calculation method is shown in the following formula: In the formula, P is the number of neighboring pixels of pixel (i, j), and m p Let f be the pixel value of the p-th pixel (i, j); the calculation method of f in the formula is shown in the following equation; u ij Let u be the pixel consistency parameter between adjacent pixels of pixel (i, j), and its calculation method is shown in the following formula. From this formula, it can be seen that when u ij When u = 0, the pixel values ​​of all neighboring pixels of (i, j) are either less than or greater than the pixel value of (i, j); when u ij When u = 2, there are several consecutive neighboring pixels whose pixel values ​​are greater than the pixel value of (i, j); when u ij When the value is greater than 2, there are non-contiguous neighboring pixels whose pixel values ​​are greater than the pixel value of (i, j). By calculating the relationship between each pixel in an image and its neighboring pixels, a relationship graph of the image can be obtained, which can effectively capture the structural information of the image. Then, the histogram corresponding to the image is obtained through the relationship graph, and the relationship value M_r of the pixels in the image is obtained. ij Given integers in the interval [1, 10], the probability of these integers occurring can be calculated to obtain the LBP histogram feature f1 of the underwater image; The extraction process of feature f2 in step S2 is as follows: B1: Perform AGGD fitting on the two obtained MSCN plots: The AGGD model is shown in the following equation, where α controls the shape of the distribution. These are scale parameters, which control the distribution on both sides of the model, and η is the mean of the distribution; in: B2: Calculate the skewness S and kurtosis K for the two MSCN plots. The specific calculation method is shown in the following formula: B3: Shape parameter α and left and right scale parameters obtained by fitting the two MSCN images using AGGD. The mean η, skewness S, and kurtosis K of the distribution are used as morphological features f2; The extraction process of moment statistic feature f3 in step S2 is as follows: The first to fourth moment statistics of the L component are calculated using the following formulas: Where, p i，j Let f represent the i-th color component of the j-th pixel in the image, and N represent the number of pixels in the image; the four moment statistics obtained are used as features f3; The extraction process of brightness histogram feature f4 in step S2 is as follows: C1: Obtain the histogram of the image from the L-component map; the value p of each pixel in the image. ij The integers are within the interval [0, 100]. C2: Calculate the probability of these integers, draw a histogram with a class interval of 10, normalize the histogram, and obtain the final brightness histogram feature f4.

2. The underwater image quality assessment method based on multi-feature fusion in color space according to claim 1, characterized in that, In the CIELab color space of step S1, the L component is used to represent the brightness of the pixel, with a value range of [0, 100], representing pure black to pure white; a represents the range from red to green, with a value range of [127, -128]; b represents the range from yellow to blue, with a value range of [127, -128].

3. The underwater image quality assessment method based on multi-feature fusion in color space according to claim 2, characterized in that, In step S1, the image is converted to the XYZ color space using the following formula. The formula for converting from XYZ color space to CIELab color space is as follows: L * =116f(Y / Y n )-16 a * =500[f(X / X n )-f(Y / Y n )] b * =200[f(Y / Y n )-f(Z / Z n )] Among them, X n Y n Z n The CIE XYZ tri-color stimulus values ​​[0.9504, 1.0000, 1.0888] were used to simulate midday sunlight, with reference white point values.

4. The underwater image quality assessment method based on multi-feature fusion in color space according to claim 1, characterized in that, The specific process of performing AGGD fitting on five color description maps and calculating kurtosis and skewness to obtain morphological feature f6 is as follows: AGGD fitting was performed on the five color description maps to determine the shape parameter α and the left and right scale parameters. The mean η of the distribution; The skewness S and kurtosis K are calculated from the five color description maps using the formula. The shape parameter α and the left and right scale parameters The distribution's mean η, skewness S, and kurtosis K are combined as characteristic f6.

Images