Quality evaluation method and system for reference-free image in low-light scene
By combining a perceptual prior feature map and a feature extraction network in a dual-branch structure, the problems of accuracy and interpretability in image quality evaluation in low-light scenes are solved, and an effective evaluation of the global quality and local distortion of low-light images is achieved.
Patent Information
- Application Number
- CN202511335308.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-18
- Publication Date
- 2025-12-19
AI Technical Summary
Existing no-reference image quality assessment methods struggle to accurately identify image degradation types and reflect human visual perception characteristics in low-light scenes, leading to discrepancies between assessment results and subjective perception.
It employs a perceptual prior feature map, a feature extraction network, a low-level parameter sharing and high-level bifurcation design, and a dual-branch structure of a quality prototype library and an artifact prototype library. Combined with feature enhancement driven by human visual system perception, it achieves accurate prediction and interpretable analysis of image quality through feature extraction and similarity calculation.
It enhances the ability to capture global quality features and local distortion features of low-light scene images, improves the accuracy and interpretability of evaluation results, fits the perception characteristics of low-light scenes, and enhances the ability to distinguish between insufficient brightness, noise and artifacts.
Smart Images
Figure CN121169874A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image quality assessment technology, and more specifically to a method and system for quality assessment of low-light scenes without reference images. Background Technology
[0002] Image quality assessment is an important research direction in computer vision, aiming to simulate the human visual system and objectively predict the subjective quality of images. Existing image quality assessment methods are mainly divided into three categories: full-reference, semi-reference, and no-reference methods. Full-reference and semi-reference methods typically rely on the original clear image as a reference, evaluating quality by comparing the difference between the distorted image and the reference image. However, in practical applications, especially in low-light scenes, it is often impossible to obtain an ideal lossless reference image, thus limiting the application scope of full-reference and semi-reference methods. No-reference image quality assessment methods, by eliminating the dependence on reference images, better meet the needs of real-world environments. Traditional no-reference methods mostly rely on manually designed feature descriptors, such as feature modeling methods based on Natural Scene Statistics (NSS). However, due to the complex degradation of low-light images, such as noise enhancement, color distortion, loss of detail, and insufficient contrast, manual features can usually only model specific distortions and cannot comprehensively reflect the multi-dimensional degradation characteristics under low-light conditions.
[0003] In recent years, deep learning has provided a new solution for no-reference image quality assessment. Models such as convolutional neural networks and Transformers can automatically extract multi-level features within an end-to-end framework. However, image degradation in low-light environments involves not only local noise and texture destruction but also overall insufficient brightness and perceptual discomfort. These complex degradation patterns make it difficult for single-feature learning mechanisms to simultaneously capture both local details and global structure. Furthermore, existing methods generally ignore the differences in human visual perception characteristics in low-light scenes, such as enhanced sensitivity to detail and increased sensitivity to brightness changes, leading to discrepancies between evaluation results and actual subjective perception. Moreover, existing methods still fall short in accurately identifying different types of degradation and reasonably modeling human perception in light of the complexity of low-light image quality degradation. On the one hand, models lack the ability to effectively distinguish distortions specific to low-light scenes; on the other hand, networks often ignore changes in human visual perception regarding detail, contrast, and color sensitivity under low-light conditions, resulting in a gap between evaluation results and subjective perception.
[0004] Therefore, how to achieve accurate prediction and interpretable analysis of image quality in low-light scenes is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides a quality assessment method and system for low-light scenes without reference images. By combining perceptual prior feature maps, feature extraction networks, low-level parameter sharing and high-level bifurcation design, a dual-branch structure of quality prototype library and artifact prototype library, and feature enhancement driven by human visual system perception, it achieves accurate prediction and interpretable analysis of image quality in low-light scenes.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A method for quality assessment of no-reference images in low-light scenes includes:
[0008] Obtain the target image to be predicted;
[0009] A perceptual prior feature map set is obtained based on the target image;
[0010] A comprehensive feature map is obtained by concatenating the prior perceptual feature map with the target image;
[0011] Based on the comprehensive feature map, feature extraction is performed to obtain global quality features and local distortion features;
[0012] Based on the global quality features and the local distortion features, the quality similarity and distortion intensity are obtained accordingly.
[0013] A quality score is obtained based on the quality similarity and the distortion intensity to evaluate the image quality of the target image.
[0014] In one embodiment, the method for obtaining the perceptual prior feature map set is as follows:
[0015] Based on the target image, brightness adaptation and contrast sensitivity calculations are performed respectively to obtain brightness adaptation feature map and contrast sensitivity feature map;
[0016] A color difference visibility feature map is obtained based on the brightness adaptive feature map;
[0017] The brightness adaptive feature map, the contrast sensitivity feature map, and the color difference visibility feature map together constitute the perceptual prior feature map set.
[0018] In one embodiment, the brightness adaptive feature map T lum (x) specifically refers to:
[0019]
[0020] L a (x)=blur(L(x) ;σ a );
[0021] Where L(x) represents the luminance component of the target image mapped to the linear color space, γ represents the coefficient controlling the shape of the response curve, and L... a (x) represents the smoothed local adaptive brightness, ε represents the stability term, blur represents the large-scale smoothing operation, and σ a Indicates the smoothing scale.
[0022] In one embodiment, the contrast sensitivity feature map T csf (x) specifically refers to:
[0023]
[0024] Among them, w f G represents the frequency band weight. f,θ M represents a bandpass filter kernel at frequency f and direction θ. f,θ (x) represents the masking term calculated based on spatial frequency and direction, k represents the masking bias, |Ω| represents the number of pixels in the window, Ω(x) represents the local window centered at x, and μ represents the element in Ω(x).
[0025] In one embodiment, the color difference visibility feature map T col (x) specifically refers to:
[0026]
[0027] Where △E(x) represents local color difference, and c1 and c2 are both coefficients that control the modulation intensity.
[0028] In one embodiment, the method for obtaining the global quality features and the local distortion features is as follows:
[0029] Based on the comprehensive feature map input to the bottom shared structure of the feature extraction network, general features are extracted;
[0030] Based on the general features, they are respectively input into the quality branch and the artifact branch of the feature extraction network. The general features are then processed based on the corresponding branch feature extraction functions to obtain the global quality features and the local distortion features.
[0031] In one embodiment, the method for obtaining the quality similarity and the distortion intensity is specifically as follows:
[0032] The similarity between the global quality features and the high-quality prototype set is calculated, and the first similarity is obtained as the quality similarity.
[0033] The similarity between the local distortion features and the distortion prototypes in the artifact prototype set is calculated to obtain a second similarity as the distortion intensity.
[0034] In one embodiment, the mass fraction Q s Specifically:
[0035]
[0036] Among them, β, and All of these represent learnable parameters. Indicates quality similarity. Indicates the intensity of distortion.
[0037] In one embodiment, it also includes:
[0038] Construct a total loss function and optimize the feature extraction network to obtain a trained feature extraction network for feature extraction;
[0039] The total loss function is composed of quality regression loss, prototype regularization loss, and branch complementarity loss.
[0040] A quality assessment system for low-light scenes without reference images includes: a feature map acquisition module, a feature map stitching module, a feature extraction module, a similarity acquisition module, and a quality assessment module;
[0041] The feature map acquisition module is used to acquire the target image to be predicted; and to acquire a perceptual prior feature map based on the target image.
[0042] The feature map stitching module is used to stitch the perceptual prior feature map set with the target image to obtain a comprehensive feature map.
[0043] The feature extraction module is used to extract features based on the comprehensive feature map to obtain global quality features and local distortion features;
[0044] The similarity acquisition module is used to acquire quality similarity and distortion intensity based on the global quality features and the local distortion features.
[0045] The quality evaluation module is used to obtain a quality score based on the quality similarity and the distortion intensity to evaluate the image quality of the target image.
[0046] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for quality assessment of low-light scenes without reference images, which has the following beneficial effects:
[0047] 1. This invention adopts a dual-branch feature extraction network structure with shared bottom-level parameters. At the high-level, features are extracted for high-quality mode and distortion mode respectively, resulting in global quality features and local distortion features. This enables the network to capture both global quality features and local distortion features in low-light images, enhancing the complementary representation ability of overall perception and detail anomalies.
[0048] 2. Alignment quality prototype library and artifact prototype library, focusing on modeling global quality mode and local distortion mode respectively, giving higher sensitivity to the similarity of dark area features to match the perception characteristics of low light scene, and enhancing the discriminability and interpretability of insufficient brightness, noise and artifacts in low light scene.
[0049] 3. This invention employs prototype regularization loss to minimize the similarity between similar prototypes and maximize the distance between dissimilar prototypes, thereby preventing prototype collapse and enhancing separability. Attached Figure Description
[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0051] Figure 1 The flowchart of a quality assessment method for low-light scenes without reference images provided by the present invention is shown.
[0052] Figure 2 This is a schematic diagram of the image quality evaluation model structure provided by the present invention.
[0053] Figure 3 This is a schematic diagram of a quality assessment system for low-light scenes without reference images, provided by the present invention. Detailed Implementation
[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0055] Example 1
[0056] like Figures 1-2As shown, this embodiment of the invention discloses a quality assessment method for low-light scenes without reference images, including the following steps. For ease of description, these steps are numbered S1 to S6, and these numbers are not intended to limit the sequential relationship between the various steps of this invention:
[0057] S1 acquires the target image to be predicted.
[0058] Furthermore, the image dataset to be predicted is obtained, and the images in the dataset are uniformly adjusted to a preset size and used as the target image set for subsequent processing.
[0059] S2 obtains a perceptual prior feature map based on the target image.
[0060] Furthermore, the method for obtaining the perceptual prior feature map is as follows:
[0061] Brightness adaptation and contrast sensitivity calculations are performed on the target image to obtain corresponding brightness adaptation feature maps and contrast sensitivity feature maps.
[0062] Color difference visibility feature map is obtained based on brightness adaptive feature map;
[0063] The brightness adaptive feature map, contrast sensitivity feature map, and color difference visibility feature map together form the perceptual prior feature map set.
[0064] Furthermore, the brightness adaptive feature map T lum (x) specifically refers to:
[0065]
[0066] L a (x)=blur(L(x) ;σ a );
[0067] Where L(x) represents the luminance component of the target image mapped to the linear color space, γ represents the coefficient controlling the shape of the response curve, and L... a (x) represents the smoothed local adaptive brightness, ε represents the stability term, blur represents the large-scale smoothing operation, and σ a This represents the smoothness scale; a larger value indicates a wider range of applicability.
[0068] Furthermore, the contrast sensitivity feature map T csf (x) specifically refers to:
[0069]
[0070] Among them, w f G represents the frequency band weight. f,θ M represents a bandpass filter kernel at frequency f and direction θ. f,θ(x) represents the masking term calculated based on spatial frequency and direction, k represents the masking bias, |Ω| represents the number of pixels in the window, Ω(x) represents the local window centered at x, and μ represents the element in Ω(x).
[0071] Furthermore, the color difference visibility feature map T col (x) specifically refers to:
[0072]
[0073] Where △E(x) represents local color difference, and c1 and c2 are both coefficients that control the modulation intensity.
[0074] S3 obtains a comprehensive feature map by stitching together the perceptual prior feature map set and the target image.
[0075] Furthermore, based on the brightness adaptive feature map T lum (x), Contrast sensitivity feature map T csf (x), Color difference visibility feature map T col (x) and target image I dist By fusing and stitching, we obtain the comprehensive feature map P:
[0076] P = concat(T) lum (x),T csf (x),T col (x),I dist ).
[0077] S4 extracts features based on the comprehensive feature map to obtain global quality features and local distortion features.
[0078] Furthermore, the methods for obtaining global quality features and local distortion features are as follows:
[0079] General features are extracted by inputting the comprehensive feature map into the underlying shared structure of the feature extraction network;
[0080] Based on the general features, they are input into the quality branch and artifact branch of the feature extraction network, respectively. The general features are then processed based on the corresponding branch feature extraction functions to obtain the global quality features and local distortion features.
[0081] Furthermore, the general feature f T Specifically:
[0082] f T =f share (P);
[0083] Among them, f share This represents the underlying shared structure of the feature extraction network.
[0084] Furthermore, the global quality feature Fg is specifically as follows:
[0085] F g =f g (f T );
[0086] Among them, f g This represents the high-level feature extraction function for the quality branch.
[0087] Furthermore, the local distortion feature F a Specifically:
[0088] F a =f a (f T );
[0089] Among them, f a This represents the high-level feature extraction function for the artifact branch.
[0090] Furthermore, the quality branch uses the high-level structure of the feature extraction network, while the artifact branch enhances the perception of local anomalies and artifact patterns based on the high-level structure of the feature extraction network. The two branches focus on acquiring features of global quality patterns and local distortion patterns, respectively.
[0091] S5 obtains quality similarity and distortion intensity based on the correspondence between global quality features and local distortion features.
[0092] Furthermore, the methods for obtaining quality similarity and distortion intensity are as follows:
[0093] The similarity is calculated based on global quality features and a set of high-quality prototypes, and the first similarity is used as the quality similarity.
[0094] The second similarity is obtained as the distortion intensity based on the similarity between the local distortion features and the distortion prototypes in the artifact prototype set.
[0095] Furthermore, quality similarity Specifically:
[0096]
[0097] Among them, F g Represents the global quality features of the target image. τ represents a high-quality prototype library containing k high-quality prototypes. g This represents a learnable temperature parameter.
[0098] Furthermore, high-quality prototypes are a set of reference features of high-perception quality pattern images. If the target image to be predicted is similar to these prototypes, it means that it is closer to a good perception pattern. Multiple high-quality prototypes form a high-quality prototype library. The high-quality prototypes are set as trainable parameters and optimized end-to-end together with the feature extraction network.
[0099] Furthermore, distortion intensity Specifically:
[0100]
[0101] Among them, F a Represents the local distortion features of the target image. τ represents a library of artifact prototypes containing k artifact prototypes. a This represents a learnable temperature parameter.
[0102] Furthermore, artifact prototypes are typical feature anchors of bad images with common distortion patterns. The higher the similarity between the target image to be predicted and these prototypes, the more it deviates from good quality. Multiple artifact prototypes form an artifact prototype library. The artifact prototypes are set as trainable parameters and optimized end-to-end together with the feature extraction network.
[0103] S6 uses quality similarity and distortion intensity to obtain a quality score to evaluate the image quality of the target image.
[0104] Furthermore, the quality fraction Q s Specifically:
[0105]
[0106] Among them, β, and All of these represent learnable parameters. Indicates quality similarity. Indicates the intensity of distortion.
[0107] Furthermore, evaluating the image quality of a target image based on a quality score specifically includes:
[0108] The obtained mass fraction Q s Normalization is performed to obtain interval scores [0,1], allowing for direct comparison between different images; if Q s If the value is greater than the set threshold, it indicates that the image is more similar to the high-quality prototype and farther away from the artifact prototype, and the predicted image quality is better; if Q s If the value is less than or equal to the set threshold, it indicates that the distortion intensity is dominant and the image quality is poor.
[0109] Furthermore, it also includes:
[0110] Construct a total loss function and optimize the training of the feature extraction network to obtain a trained feature extraction network for feature extraction;
[0111] The total loss function consists of the quality regression loss, the prototype regularization loss, and the branch complementarity loss.
[0112] Furthermore, during training, the quality regression loss L is calculated. reg Used to fit the mapping relationship between predicted scores and subjective quality scores:
[0113]
[0114] Among them, y i '、y i These represent the predicted score and the corresponding MOS value, respectively, and N represents the number of samples.
[0115] Furthermore, during training, the prototype regularization loss L is calculated. pro Minimize the similarity between similar prototypes and maximize the distance between dissimilar prototypes to ensure the discriminativeness and diversity of prototype distribution:
[0116]
[0117] in, and All of these represent prototypes from a high-quality prototype library. and Both represent prototypes in the artifact prototype library, and λ>0 controls the cross-library interval strength.
[0118] Furthermore, during training, the branch complementarity loss L is calculated. comp Suppressing high-quality branches and artifact branches to learn redundant feature distributions:
[0119]
[0120] Among them, F i g F represents the global quality feature of the i-th target image. i a This represents the local distortion features of the i-th target image.
[0121] Furthermore, by jointly optimizing the quality regression loss, prototype regularization loss, and branch complementarity loss, a total loss function is obtained to ensure prediction accuracy:
[0122] L total =ω1L reg +ω2L pro +ω3L comp ;
[0123] Here, ω1, ω2, and ω3 all represent the corresponding weights.
[0124] Example 2
[0125] like Figure 3 As shown, based on the same inventive concept, the present invention also provides a quality assessment system for low-light scenes without reference images, including: a feature map acquisition module, a feature map stitching module, a feature extraction module, a similarity acquisition module, and a quality assessment module;
[0126] The feature map acquisition module is used to acquire the target image to be predicted; and to acquire a perceptual prior feature map based on the target image.
[0127] The feature map stitching module is used to stitch together the target image with the perceptual prior feature map set to obtain a comprehensive feature map;
[0128] The feature extraction module is used to extract features based on the comprehensive feature map to obtain global quality features and local distortion features;
[0129] The similarity acquisition module is used to obtain quality similarity and distortion intensity based on the correspondence between global quality features and local distortion features;
[0130] The quality assessment module is used to obtain a quality score based on quality similarity and distortion intensity to evaluate the image quality of the target image.
[0131] Furthermore, in this embodiment, the functional implementation methods of each functional module correspond one-to-one with the methods described above, and will not be repeated here.
[0132] Example 3
[0133] Based on the same inventive concept, the present invention also provides an electronic device, which includes a processor and a memory, wherein the memory stores instructions, characterized in that the instructions are loaded and executed by the processor to implement a quality evaluation method for a low-light scene without reference image as described in Embodiment 1.
[0134] Based on the same inventive concept, the present invention also provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;
[0135] Memory, used to store computer programs;
[0136] When the processor executes a program stored in the memory, it is able to implement a quality assessment method for a low-light scene without a reference image, as described in Example 1.
[0137] The electronic device may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can call logical instructions in the memory to execute a hardware implementation method of an on-ring polynomial multiplier as described in Embodiment 1.
[0138] Furthermore, when the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0139] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0140] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for quality assessment of no-reference images in low-light scenes, characterized in that, include: Obtain the target image to be predicted; A perceptual prior feature map set is obtained based on the target image; A comprehensive feature map is obtained by concatenating the prior perceptual feature map with the target image; Based on the comprehensive feature map, feature extraction is performed to obtain global quality features and local distortion features; Based on the global quality features and the local distortion features, the quality similarity and distortion intensity are obtained accordingly. A quality score is obtained based on the quality similarity and the distortion intensity to evaluate the image quality of the target image.
2. The quality assessment method for no-reference images in low-light scenes according to claim 1, characterized in that, The method for obtaining the prior perceptual feature map set is as follows: Based on the target image, brightness adaptation and contrast sensitivity calculations are performed respectively to obtain brightness adaptation feature map and contrast sensitivity feature map; A color difference visibility feature map is obtained based on the brightness adaptive feature map; The brightness adaptive feature map, the contrast sensitivity feature map, and the color difference visibility feature map together constitute the perceptual prior feature map set.
3. The method for quality assessment of referenceless images in low-light scenes according to claim 2, characterized in that, The brightness adaptive feature map T lum (x) specifically refers to: L a (x)=blur(L(x) σ a ): Where L(x) represents the luminance component of the target image mapped to the linear color space, γ represents the coefficient controlling the shape of the response curve, and L... a (x) represents the smoothed local adaptive brightness, ε represents the stability term, blur represents the large-scale smoothing operation, and σ a Indicates the smoothing scale.
4. The quality assessment method for no-reference images in low-light scenes according to claim 3, characterized in that, The contrast sensitivity feature map T csf (x) specifically refers to: Among them, w f G represents the frequency band weight. f,θ M represents a bandpass filter kernel at frequency f and direction θ. f,θ (x) represents the masking term calculated based on spatial frequency and direction, k represents the masking bias, |Ω| represents the number of pixels in the window, Ω(x) represents the local window centered at x, and μ represents the element in Ω(x).
5. A method for quality assessment of referenceless images in low-light scenes according to claim 3, characterized in that, Color difference visibility feature map T col (x) specifically refers to: Where △E(x) represents local color difference, and c1 and c2 are both coefficients that control the modulation intensity.
6. The method for quality assessment of no-reference images in low-light scenes according to claim 1, characterized in that, The method for obtaining the global quality features and the local distortion features is as follows: Based on the comprehensive feature map input to the bottom shared structure of the feature extraction network, general features are extracted; Based on the general features, they are respectively input into the quality branch and the artifact branch of the feature extraction network. The general features are then processed based on the corresponding branch feature extraction functions to obtain the global quality features and the local distortion features.
7. The method for quality assessment of referenceless images in low-light scenes according to claim 1, characterized in that, The methods for obtaining the quality similarity and the distortion intensity are as follows: The similarity between the global quality features and the high-quality prototype set is calculated, and the first similarity is obtained as the quality similarity. The similarity between the local distortion features and the distortion prototypes in the artifact prototype set is calculated to obtain a second similarity as the distortion intensity.
8. The method for quality assessment of referenceless images in low-light scenes according to claim 1, characterized in that, The mass fraction Q s Specifically: Among them, β, and All of these represent learnable parameters. Indicates quality similarity. Indicates the intensity of distortion.
9. A method for quality assessment of referenceless images in low-light scenes according to claim 6, characterized in that, Also includes: Construct a total loss function and optimize the feature extraction network to obtain a trained feature extraction network for feature extraction; The total loss function is composed of quality regression loss, prototype regularization loss, and branch complementarity loss.
10. A quality assessment system for no-reference images in low-light scenes, used to perform a quality assessment method for no-reference images in low-light scenes as described in any one of claims 1-9, characterized in that, include: The module includes a feature map acquisition module, a feature map stitching module, a feature extraction module, a similarity acquisition module, and a quality evaluation module. The feature map acquisition module is used to acquire the target image to be predicted; and to acquire a perceptual prior feature map based on the target image. The feature map stitching module is used to stitch the perceptual prior feature map set with the target image to obtain a comprehensive feature map. The feature extraction module is used to extract features based on the comprehensive feature map to obtain global quality features and local distortion features; The similarity acquisition module is used to acquire quality similarity and distortion intensity based on the global quality features and the local distortion features. The quality evaluation module is used to obtain a quality score based on the quality similarity and the distortion intensity to evaluate the image quality of the target image.