An image quality evaluation method and system based on dual-domain distribution difference modeling
By performing distribution modeling in the spatial and frequency domains, and combining multi-scale feature extraction and integral probability measurement, the accuracy and consistency problems of existing image quality assessment methods under complex distortion conditions are solved, achieving more efficient image quality assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-29
AI Technical Summary
Existing full-reference image quality assessment methods struggle to accurately characterize perceptual differences under complex distortion conditions and lack a unified model of the distortion generation process and human visual perception process, resulting in inconsistencies between the assessment results and human subjective perception, and limited generalization ability.
A method based on dual-domain distribution difference modeling is adopted, which models the distribution of distorted images and reference images in the spatial domain and frequency domain respectively. Multi-scale distribution differences of images are calculated by multi-scale feature extraction and integral probability measurement, and weighted fusion is performed based on perceptual weights to generate the final image quality evaluation result.
It improves the accuracy and stability of image quality assessment, enhances the consistency between assessment results and human subjective perception, and has good generalization ability and cross-scene adaptability.
Smart Images
Figure CN122115355A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision and image quality assessment technology, and more specifically to an image quality assessment method and system based on dual-domain distribution difference modeling. Background Technology
[0002] Image Quality Assessment (IQA) is an important research direction in computer vision and image processing. Its purpose is to objectively and quantitatively evaluate the visual quality of images, ensuring that the evaluation results are as consistent as possible with human visual perception. IQA results are widely used in image compression, transmission, and storage, and are of great significance for performance optimization and algorithm decision-making in related systems. Based on whether they rely on the original reference image, existing image quality assessment methods are generally divided into three categories: full-reference, semi-reference, and no-reference. Among them, full-reference image quality assessment methods can utilize both the distorted image and its corresponding reference image during the evaluation process, providing the most comprehensive information. Therefore, they have significant advantages in evaluation accuracy and stability, and are a commonly used evaluation method in high-precision vision systems.
[0003] Traditional full-reference image quality assessment methods are mostly based on assumptions of pixel differences or structural similarity, estimating image quality by comparing artificially designed features such as brightness, contrast, and structure. While simple to implement, these methods struggle to accurately characterize perceptual differences under complex distortion conditions, often resulting in discrepancies between assessment results and human subjective perception. With the development of deep learning, full-reference image quality assessment methods based on deep neural networks have emerged, improving assessment performance to some extent by extracting multi-level features and measuring differences in the feature space. However, these methods typically rely on specific network structures and training data, limiting their generalization ability to unknown distortion types or cross-scene applications, and lacking interpretable modeling of quality degradation mechanisms. In recent years, some studies have attempted to assess image quality from a statistical distribution perspective, reflecting the degree of quality degradation by comparing the distribution differences between distorted and reference images in the feature space. However, existing methods remain relatively simplistic in their distribution representation and distance metric design, failing to simultaneously characterize the multi-scale differences in the distortion generation process and human perception, thus limiting assessment accuracy and stability.
[0004] In existing related technologies, some full-reference image quality assessment methods introduce local region segmentation, masking mechanisms, or attention models. By assigning different weights to different spatial regions or local features in the image, they enhance the model's ability to focus on key regions, thereby improving the consistency between the quality assessment results and human visual perception. However, these methods typically require explicit local region segmentation of the image and rely on specific masking strategies, feature aggregation methods, and network structure designs. Their quality assessment process essentially remains at the level of local feature selection and feature weighting, making it difficult to characterize the distributional differences between the distorted image and the reference image from an overall statistical perspective. Furthermore, these methods often focus on spatial domain feature modeling, insufficiently considering the frequency domain statistical characteristics inherent in the distortion generation process, and failing to coordinate the relationship between the distortion generation mechanism and the human visual perception mechanism within a unified framework. This results in limited stability and generalization ability of the evaluation results under complex distortion types or cross-scenario application conditions.
[0005] Therefore, there is an urgent need for an image quality assessment method that can make full use of reference image information, jointly model distortion generation and distortion perception at the distribution level, and does not rely on specific artificial features and data training, so as to improve the accuracy, generalization and reliability of the assessment results. Summary of the Invention
[0006] In view of the above problems, the present invention proposes an image quality evaluation method and system based on dual-domain distribution difference modeling to overcome or at least partially solve the above problems.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] In a first aspect, the present invention provides an image quality assessment method based on dual-domain distribution difference modeling, comprising: The distorted image to be evaluated and the corresponding reference image are obtained, and the distorted image and the reference image are preprocessed. Distribution modeling is performed on the preprocessed distorted image and the reference image in the spatial domain and the frequency domain, respectively, to obtain the spatial domain distribution characteristics and the frequency domain distribution characteristics of the distorted image and the reference image. Multi-scale feature extraction is performed on the spatial domain distribution features and frequency domain distribution features respectively to obtain spatial domain distribution embedding representations and frequency domain distribution embedding representations of the distorted image and the reference image at different scale levels; The spatial domain distribution embedding representation and the frequency domain distribution embedding representation are calculated based on the integral probability metric to obtain the multi-scale spatial domain distribution difference and multi-scale frequency domain distribution difference between the distorted image and the reference image. Based on the multi-scale spatial domain distribution differences, the perceptual weights corresponding to the frequency domain distribution differences at each scale are determined, and the multi-scale frequency domain distribution differences are weighted and fused based on the perceptual weights to generate the final image quality evaluation result.
[0009] Furthermore, the preprocessing includes one or more operations among size adjustment, pixel value normalization, format conversion, channel rearrangement, downsampling, or scale normalization.
[0010] Furthermore, the step of performing distribution modeling on the preprocessed distorted image and the reference image in the spatial and frequency domains respectively to obtain the spatial and frequency domain distribution features of the distorted image and the reference image specifically includes: In the spatial domain, based on the pixel values or pixel features of the preprocessed distorted image and the reference image, a spatial domain statistical distribution reflecting the image structure and human visual perception characteristics is constructed as a spatial domain distribution feature. In the frequency domain, frequency domain transformation is performed on the preprocessed distorted image and the reference image respectively to construct a multi-band frequency domain distribution that reflects the characteristics of the distortion spectrum change, which serves as the frequency domain distribution feature.
[0011] Furthermore, the frequency domain transformation process includes at least one of wavelet transform, Fourier transform, or discrete cosine transform.
[0012] Furthermore, the multi-scale feature extraction is achieved through a pre-trained feature extraction network, and the spatial domain distribution embedding representation and the frequency domain distribution embedding representation are obtained by extracting features from multiple different levels of the feature extraction network.
[0013] Furthermore: The multi-scale spatial domain distribution difference is measured by calculating the difference between the second-order statistical matrices of the spatial domain distribution embedding representations of the distorted image and the reference image at each scale level. The multi-scale frequency domain distribution difference is measured by calculating the difference between the statistical centers of the frequency domain distribution embedding representations corresponding to the distorted image and the reference image at each scale level.
[0014] Furthermore, the perception weights are obtained by normalizing the differences in the multi-scale spatial domain distribution.
[0015] Secondly, the present invention provides an image quality assessment system based on dual-domain distribution difference modeling. Applying the aforementioned image quality assessment method based on dual-domain distribution difference modeling, the system includes: The acquisition and preprocessing module is used to acquire the distorted image to be evaluated and the corresponding reference image, and to preprocess the distorted image and the reference image. The spatial domain distribution modeling module is used to perform distribution modeling on the preprocessed distorted image and the reference image in the spatial domain to obtain the spatial domain distribution features of the distorted image and the reference image. The frequency domain distribution modeling module is used to perform distribution modeling on the preprocessed distorted image and the reference image in the frequency domain to obtain the frequency domain distribution characteristics of the distorted image and the reference image. The multi-scale feature extraction module is used to extract multi-scale features from the spatial domain distribution features and the frequency domain distribution features respectively, so as to obtain spatial domain distribution embedding representations and frequency domain distribution embedding representations of the distorted image and the reference image at different scale levels. The distribution difference calculation module is used to calculate the spatial domain distribution embedding representation and the frequency domain distribution embedding representation based on the integral probability metric, so as to obtain the multi-scale spatial domain distribution difference and multi-scale frequency domain distribution difference between the distorted image and the reference image. The quality scoring fusion module is used to determine the perceptual weights corresponding to the frequency domain distribution differences at each scale based on the multi-scale spatial domain distribution differences, and to perform weighted fusion of the multi-scale frequency domain distribution differences based on the perceptual weights to generate the final image quality evaluation result.
[0016] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described image quality evaluation method based on dual-domain distribution difference modeling.
[0017] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described image quality evaluation method based on dual-domain distribution difference modeling.
[0018] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an image quality evaluation method based on dual-domain distribution difference modeling, which has the following beneficial effects: This invention employs an image quality assessment method based on distribution differences, which compares the overall distribution of distorted images with that of reference images. This avoids the information loss problem caused by relying solely on local features or manually designed features, and effectively improves the accuracy and stability of image quality assessment results.
[0019] This invention constructs distribution models in the spatial and frequency domains respectively, and jointly models the multi-scale distribution differences based on integral probability measures, thereby achieving a collaborative characterization of the image distortion generation process and the human eye perception process, and thus realizing a more comprehensive and reasonable image quality representation.
[0020] This invention utilizes the differences in multi-scale spatial domain distribution as the weighting basis for the differences in multi-scale frequency domain distribution, making the contribution of distortion at different scales to the final quality evaluation result more consistent with the subjective perception of the human eye, thereby further improving the consistency between the evaluation result and the subjective quality score. Attached Figure Description
[0021] 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.
[0022] Figure 1 This is a schematic diagram of the image quality evaluation method based on dual-domain distribution difference modeling provided in this embodiment of the invention.
[0023] Figure 2 This is a schematic diagram of the image quality evaluation method framework based on dual-domain distribution difference modeling provided in an embodiment of the present invention. Detailed Implementation
[0024] 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.
[0025] This invention discloses an image quality assessment method based on dual-domain distribution difference modeling, such as... Figure 1 and Figure 2 As shown, it includes: S1. Obtain the distorted image to be evaluated and the corresponding reference image, and preprocess the distorted image and the reference image; S2. Perform distribution modeling on the preprocessed distorted image and the reference image in the spatial domain and frequency domain respectively to obtain the spatial domain distribution characteristics and frequency domain distribution characteristics of the distorted image and the reference image. S3. Perform multi-scale feature extraction on spatial domain distribution features and frequency domain distribution features respectively to obtain spatial domain distribution embedding representations and frequency domain distribution embedding representations of distorted images and reference images at different scale levels; S4. Calculate the spatial domain distribution embedding representation and the frequency domain distribution embedding representation based on the integral probability metric to obtain the multi-scale spatial domain distribution difference and multi-scale frequency domain distribution difference between the distorted image and the reference image. S5. Based on the differences in spatial domain distribution at multiple scales, determine the perceptual weights corresponding to the differences in frequency domain distribution at each scale, and perform weighted fusion of the differences in frequency domain distribution at multiple scales based on the perceptual weights to generate the final image quality evaluation result.
[0026] Next, each of the above steps will be explained in detail.
[0027] In step S1 above, the distorted image to be evaluated and its corresponding reference image are obtained, and the distorted image and reference image are preprocessed to unify the image size, eliminate differences in resolution and numerical distribution of the input images, and provide consistent input conditions for subsequent distribution modeling; specifically, the above preprocessing of the distorted image and reference image includes one or more of the following operations: (1) Resize the distorted image and the reference image to make them have the same spatial resolution; (2) Normalize the pixel values of the distorted image and the reference image to eliminate the differences in the numerical range between the different distorted images and the reference image; (3) Convert the format or rearrange the channels of the distorted image and the reference image to meet the input requirements of subsequent distribution modeling and feature extraction.
[0028] (4) Downsampling or scale normalization is performed on the distorted image and the reference image to reduce computational complexity and ensure the consistency of distribution modeling results under different image content and resolution conditions.
[0029] The above preprocessing operations are performed without introducing additional distortion or changing the relative quality relationship of the images, so as to ensure that the subsequent distribution difference calculation results can truly reflect the degree of quality degradation of the distorted image relative to the reference image.
[0030] In step S2 above, distribution modeling is performed on the preprocessed distorted image and the reference image in the spatial domain and frequency domain, respectively, to obtain the spatial domain distribution features and frequency domain distribution features of the distorted image and the reference image; specifically including: (1) Based on the preprocessed distorted image and the reference image, construct the corresponding spatial domain distribution representation: In the spatial domain, statistical modeling is performed based on the pixel values or pixel features of the preprocessed distorted image and the reference image, respectively, to obtain the spatial domain statistical distribution that reflects the overall image structure information and human visual perception characteristics, which serves as the spatial domain distribution feature; its form is expressed as:
[0031] in, Represents the spatial domain distribution characteristics of distorted images; X represents the pixel value or pixel feature of the distorted image; X represents the spatial domain feature space. This represents the spatial domain distribution characteristics of the reference image; Represents the pixel values or pixel features of the reference image; (2) Based on the preprocessed distorted image and the reference image, construct the corresponding frequency domain distribution features: In the frequency domain, frequency domain transformation is performed on both the preprocessed distorted image and the reference image, mapping them from the spatial domain to the frequency domain representation. This constructs a multi-band frequency domain distribution that reflects the spectral variation characteristics of the distortion, serving as the frequency domain distribution feature; its form is as follows:
[0032] in, Represents the frequency domain distribution characteristics of distorted images; This represents the frequency domain distribution characteristics of the reference image; This represents a frequency domain transformation operator used to map a spatial domain signal to a spectral representation.
[0033] The aforementioned frequency domain transformation process includes at least one of wavelet transform, Fourier transform, or discrete cosine transform, and concatenates the low-frequency approximation coefficients with the high-frequency detail coefficients as the input representation for frequency domain distribution modeling. The frequency domain distribution consists of multiple frequency levels or sub-bands, and its representation is as follows:
[0034] in, K This represents the total number of frequency levels or sub-bands in the frequency domain distribution. Different frequency levels or sub-bands are used to distinguish between low-frequency structural information and high-frequency detail information in an image, thus providing a basis for subsequent multi-scale feature extraction and distribution difference calculation. Indicates the first Frequency domain distribution characteristics of distorted images in a frequency level or sub-band; Indicates the first Frequency domain distribution characteristics of reference images in a frequency level or sub-band.
[0035] In step S3 above, multi-scale feature extraction is performed on the spatial domain distribution features and frequency domain distribution features respectively to obtain the spatial domain distribution embedding representation and frequency domain distribution embedding representation of the distorted image and the reference image at different scale levels. By using a pre-trained feature extraction network to extract multi-scale features from spatial and frequency domain distribution features, the original distribution representation is mapped to a high-dimensional feature space to obtain distribution embedding representations at different levels, thereby characterizing the distribution characteristics of images at different perceptual scales.
[0036] To achieve multi-scale distribution modeling, spatial domain distribution embedding representations and frequency domain distribution embedding representations are extracted from multiple layers of the feature extraction network to form multi-scale distribution embeddings. Distribution embeddings at different levels are used to characterize the distribution characteristics of images at different scales, from low-level structural information to high-level semantic information, providing a multi-scale foundation for subsequent distribution difference calculation and quality score fusion.
[0037] The aforementioned pre-trained feature extraction network can be a pre-trained VGG network.
[0038] In step S4 above, in order to theoretically characterize the difference between the two distributions, an integral probability metric (IPM) is introduced to calculate the spatial domain distribution embedding representation and the frequency domain distribution embedding representation, so as to obtain the distribution difference between the distorted image and the reference image in the spatial domain and the frequency domain, respectively, as the multi-scale spatial domain distribution difference and the multi-scale frequency domain distribution difference. The multi-scale spatial domain distribution difference and the multi-scale frequency domain distribution difference are the distribution difference measurement results in different domains and different scales, which are used to provide a basis for subsequent weight determination and quality score fusion.
[0039] The basic form of the above integral probability measure is expressed as:
[0040] in, P and Q Indicates the two distributions to be compared; F This represents the function space that satisfies the preset constraints. Candidate functions in the function space; Let IPM be the distance in the function space. F Below, distribution P and Q The degree of difference between them; Indicates distribution P Below, function Expected value; Indicates distribution Q Below, function Expected value; The supremum represents the summum of all possible functions. In the given information, select the value corresponding to the function that maximizes the absolute value of the difference between the two expectations; In candidate functions Given that it can be represented as the inner product of feature maps, the integral probability metric can be equivalently represented as the embedding differences distributed in the feature space:
[0041] in, This represents a feature mapping function implemented by a feature extraction network, used to extract distributed embedding representations from distributed features; q It represents the norm form corresponding to the function space constraints.
[0042] The above modeling method can transform the probability distribution differences into computable distances in the feature space, providing a theoretical basis for subsequent distribution difference calculations.
[0043] In this embodiment of the invention, to further adapt to the image distortion generation mechanism and the characteristics of human visual perception, different function spaces are constructed for each scale level. i Calculate the multi-scale spatial domain distribution differences and multi-scale frequency domain distribution differences between the distorted image and the reference image; specifically: (1) Define the function space for mean embedding as follows:
[0044] in, The vector needed to compute the inner product, For network parameters, p-norm Represents the feature mapping function. It is represented in functional form; in this embodiment, the mapping process is implemented by a pre-trained feature extraction network, therefore the parameters... The parameters are fixed to those of the pre-trained network, requiring no additional training. This yields the mean-matching distance used to characterize the differences in statistical centers.
[0045] Calculate the multi-scale spatial domain distribution difference between the distorted image and the reference image. :
[0046] in, Indicates the first i Multi-scale spatial domain distribution differences between distorted images and reference images at various scales; This represents the function space used for calculating spatial domain distribution differences; Spatial domain distribution embedding representation of distorted images; The spatial domain distribution embedding representation of the reference image; the multi-scale spatial domain distribution difference Used to characterize structural differences and correlation changes related to human visual perception.
[0047] In a preferred embodiment, the above-mentioned multi-scale spatial domain distribution differences The difference can be calculated using a form based on second-order statistical properties, expressed as:
[0048] in, This represents a second-order statistical matrix constructed from feature embeddings; Denotes the matrix norm, such as the nuclear norm or its equivalent form.
[0049] (2) Define the function space corresponding to the covariance embedding to characterize the correlation and structural differences of the distribution:
[0050] in, , To satisfy the orthogonality constraint matrix, , This is the transpose of the corresponding matrix. For network parameters, Represents the feature mapping function. Represented in function form, The identity matrix is used to represent the second-order statistical structure in the feature space. This leads to the covariance matching distance, which characterizes the differences between statistical distributions.
[0051] Calculate the multi-scale frequency domain distribution difference between the distorted image and the reference image. :
[0052] in, Indicates the first i The multi-scale frequency domain distribution difference between the distorted image and the reference image at each scale; F fre This represents the function space used for calculating frequency domain distribution differences; Frequency domain distribution embedding representation of distorted images; The frequency domain distribution embedding representation of the reference image; the multi-scale frequency domain distribution difference It is used to characterize the energy changes or information loss intensity of distortion at the spectral level.
[0053] In a preferred embodiment, the above-mentioned multi-scale frequency domain distribution differences The difference can be calculated using the form of variation based on the statistical center, and expressed as:
[0054] in, (This represents the statistical center of the feature embedding, such as the mean vector). It represents the L2 norm or its equivalent form.
[0055] Furthermore, multi-scale spatial domain distribution differences Differences from multi-scale frequency domain distribution Calculations were performed at multiple levels to obtain the corresponding multi-scale spatial domain difference sets. Multiscale frequency domain difference set This is used in subsequent steps to determine weights based on spatial domain differences.
[0056] In step S5 above, based on the multi-scale spatial domain distribution differences, the perceptual weights corresponding to the frequency domain distribution differences at each scale are determined, and the multi-scale frequency domain distribution differences are weighted and fused based on the perceptual weights to generate the final image quality evaluation result. Specifically: (1) The spatial domain distribution difference is used to reflect the human eye's sensitivity to distortion at different scales. The distribution differences at each scale are normalized to obtain the perceptual weights, which are used to characterize the relative influence of frequency domain distortion at different scales on the final visual quality. The perceptual weights are expressed as:
[0057] in, Indicates the first i The perceptual weights corresponding to each scale; Indicates the first i Multi-scale spatial domain distribution differences between distorted images and reference images at various scales; Indicates the first j Multi-scale spatial domain distribution differences between distorted images and reference images at various scales; L Indicates the total number of scales.
[0058] (2) Based on perceptual weights, the differences in frequency domain distribution at multiple scales are weighted and fused to obtain a comprehensive distortion measure. D Its form is expressed as:
[0059] in, Indicates the first i The multi-scale frequency domain distribution difference between the distorted image and the reference image at each scale; (3) To enhance the numerical stability and discriminative power of the evaluation results, a comprehensive distortion measure can be applied. D By performing a monotonic mapping, the final image quality evaluation result is obtained. S :
[0060] in, This represents a monotonic mapping function, preferably a logarithmic function, a linear function, or an equivalent form thereof. This image quality assessment result... S Used to characterize the overall visual quality level of a distorted image relative to a reference image, and the image quality evaluation result S It has a high degree of consistency with human subjective perception.
[0061] Based on the same inventive concept, embodiments of the present invention also provide an image quality assessment system based on dual-domain distribution difference modeling. Applying the above-described image quality assessment method based on dual-domain distribution difference modeling, the system includes: The acquisition and preprocessing module is used to acquire the distorted image to be evaluated and the corresponding reference image, and to preprocess the distorted image and the reference image. The spatial domain distribution modeling module is used to perform distribution modeling on the preprocessed distorted image and the reference image in the spatial domain to obtain the spatial domain distribution features of the distorted image and the reference image. The frequency domain distribution modeling module is used to perform distribution modeling on the preprocessed distorted image and the reference image in the frequency domain to obtain the frequency domain distribution characteristics of the distorted image and the reference image. The multi-scale feature extraction module is used to extract multi-scale features from spatial domain distribution features and frequency domain distribution features respectively, and obtain spatial domain distribution embedding representations and frequency domain distribution embedding representations of distorted images and reference images at different scale levels. The distribution difference calculation module is used to calculate the spatial domain distribution embedding representation and the frequency domain distribution embedding representation based on the integral probability metric, so as to obtain the multi-scale spatial domain distribution difference and multi-scale frequency domain distribution difference between the distorted image and the reference image. The quality scoring fusion module is used to determine the perceptual weights corresponding to the frequency domain distribution differences at each scale based on the multi-scale spatial domain distribution differences, and to perform weighted fusion of the multi-scale frequency domain distribution differences based on the perceptual weights to generate the final image quality evaluation result.
[0062] Since the principle behind the problem solved by this system is similar to the aforementioned image quality assessment method based on dual-domain distribution difference modeling, the implementation of this system can be found in the implementation of the aforementioned method, and the repetitions will not be repeated.
[0063] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described image quality evaluation method based on dual-domain distribution difference modeling.
[0064] Based on the same inventive concept, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described image quality evaluation method based on dual-domain distribution difference modeling.
[0065] In summary, this invention provides an image quality assessment method and system based on dual-domain distribution difference modeling. It eliminates the need for training on specific datasets or distortion types, maintains consistent evaluation performance across different image contents and various distortion conditions, and exhibits good generalization and cross-scene adaptability. It is suitable for applications such as image processing, compression coding, image enhancement, and visual perception quality analysis. Furthermore, the image quality assessment process of this invention has clear statistical significance, the relationships between functional modules are clear, and the evaluation results are highly interpretable, which enhances the method's engineering usability and practical application value, making it applicable to various complex application scenarios such as image processing and visual perception analysis.
[0066] 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 systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0067] 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. An image quality assessment method based on dual-domain distribution difference modeling, characterized in that, include: The distorted image to be evaluated and the corresponding reference image are obtained, and the distorted image and the reference image are preprocessed. Distribution modeling is performed on the preprocessed distorted image and the reference image in the spatial domain and the frequency domain, respectively, to obtain the spatial domain distribution characteristics and the frequency domain distribution characteristics of the distorted image and the reference image. Multi-scale feature extraction is performed on the spatial domain distribution features and frequency domain distribution features respectively to obtain spatial domain distribution embedding representations and frequency domain distribution embedding representations of the distorted image and the reference image at different scale levels; The spatial domain distribution embedding representation and the frequency domain distribution embedding representation are calculated based on the integral probability metric to obtain the multi-scale spatial domain distribution difference and multi-scale frequency domain distribution difference between the distorted image and the reference image. Based on the multi-scale spatial domain distribution differences, the perceptual weights corresponding to the frequency domain distribution differences at each scale are determined, and the multi-scale frequency domain distribution differences are weighted and fused based on the perceptual weights to generate the final image quality evaluation result.
2. The image quality assessment method based on dual-domain distribution difference modeling as described in claim 1, characterized in that, The preprocessing includes one or more of the following operations: size adjustment, pixel value normalization, format conversion, channel rearrangement, downsampling, or scale normalization.
3. The image quality assessment method based on dual-domain distribution difference modeling as described in claim 1, characterized in that, The process of performing distribution modeling on the preprocessed distorted image and the reference image in the spatial and frequency domains respectively to obtain the spatial and frequency domain distribution features of the distorted image and the reference image specifically includes: In the spatial domain, based on the pixel values or pixel features of the preprocessed distorted image and the reference image, a spatial domain statistical distribution reflecting the image structure and human visual perception characteristics is constructed as a spatial domain distribution feature. In the frequency domain, frequency domain transformation is performed on the preprocessed distorted image and the reference image respectively to construct a multi-band frequency domain distribution that reflects the characteristics of the distortion spectrum change, which serves as the frequency domain distribution feature.
4. The image quality assessment method based on dual-domain distribution difference modeling as described in claim 3, characterized in that, The frequency domain transformation process includes at least one of wavelet transform, Fourier transform, or discrete cosine transform.
5. The image quality assessment method based on dual-domain distribution difference modeling as described in claim 1, characterized in that, The multi-scale feature extraction is achieved through a pre-trained feature extraction network. The spatial domain distribution embedding representation and the frequency domain distribution embedding representation are obtained by extracting features from multiple different levels of the feature extraction network.
6. The image quality assessment method based on dual-domain distribution difference modeling as described in claim 1, characterized in that: The multi-scale spatial domain distribution difference is measured by calculating the difference between the second-order statistical matrices of the spatial domain distribution embedding representations of the distorted image and the reference image at each scale level. The multi-scale frequency domain distribution difference is measured by calculating the difference between the statistical centers of the frequency domain distribution embedding representations corresponding to the distorted image and the reference image at each scale level.
7. The image quality assessment method based on dual-domain distribution difference modeling as described in claim 1, characterized in that, The perception weights are obtained by normalizing the differences in the distribution of the multi-scale spatial domain.
8. An image quality assessment system based on dual-domain distribution difference modeling, characterized in that, The image quality assessment method based on dual-domain distribution difference modeling as described in any one of claims 1-7, the system comprising: The acquisition and preprocessing module is used to acquire the distorted image to be evaluated and the corresponding reference image, and to preprocess the distorted image and the reference image. The spatial domain distribution modeling module is used to perform distribution modeling on the preprocessed distorted image and the reference image in the spatial domain to obtain the spatial domain distribution features of the distorted image and the reference image. The frequency domain distribution modeling module is used to perform distribution modeling on the preprocessed distorted image and the reference image in the frequency domain to obtain the frequency domain distribution characteristics of the distorted image and the reference image. The multi-scale feature extraction module is used to extract multi-scale features from the spatial domain distribution features and the frequency domain distribution features respectively, so as to obtain spatial domain distribution embedding representations and frequency domain distribution embedding representations of the distorted image and the reference image at different scale levels. The distribution difference calculation module is used to calculate the spatial domain distribution embedding representation and the frequency domain distribution embedding representation based on the integral probability metric, so as to obtain the multi-scale spatial domain distribution difference and multi-scale frequency domain distribution difference between the distorted image and the reference image. The quality scoring fusion module is used to determine the perceptual weights corresponding to the frequency domain distribution differences at each scale based on the multi-scale spatial domain distribution differences, and to perform weighted fusion of the multi-scale frequency domain distribution differences based on the perceptual weights to generate the final image quality evaluation result.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the image quality evaluation method based on dual-domain distribution difference modeling as described in any one of claims 1 to 7.
10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the image quality evaluation method based on dual-domain distribution difference modeling as described in any one of claims 1 to 7.