Full-reference image quality evaluation method and system based on wavelet transform and Hausdorff distance

Abstract

The invention relates to the technical field of image processing, and particularly provides a full-reference image quality evaluation method based on wavelet transform and Hausdorff distance, which comprises the following steps of: realizing efficient evaluation on image quality by combining a frequency domain change technology and deep learning; perceptual relevant features are extracted through wavelet transform, multi-scale perceptual information is captured, and how a human visual system analyzes visual information of various scales and directions in space and frequency domains is simulated. Secondly, the distribution similarity measurement module based on the Hausdorff distance can better evaluate the difference between the feature distribution of the reference image and the feature distribution of the distorted image. The method provided by the invention can accurately capture the perception quality difference without training data or subjective quality scores, and a large number of experiments show that the method is excellent in performance and highlights the ability of the method closely related to the perception of a human visual system.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, in particular to a full reference image quality evaluation method based on wavelet transform and Hausdorff distance. Background Art

[0002] Image quality assessment is a fundamental problem in the field of image processing and computer vision, and is widely used in image / video coding, super-resolution reconstruction, and image / video visual quality enhancement. This task aims to objectively evaluate the perceived quality of images by building a model that simulates the perception mechanism of the human visual system and measuring the overall quality of the image in a digital way.

[0003] Full-reference image quality assessment estimates the quality score of the distorted image by comparing the difference in human eye perception between two images. This algorithm can be used to evaluate and guide image compression coding, as a test benchmark for image processing algorithms, and as an optimization target for image processing tasks.

[0004] Existing full-reference image quality assessment methods mainly rely on manually annotated subjective quality scores and feature extraction based on deep learning. However, these methods have weak generalization capabilities when faced with diverse distortion types and unseen images, and do not adequately model the frequency dependence of distortion types. In addition, point-to-point feature metrics often ignore the global information of feature distribution, limiting the accuracy and robustness of the evaluation results. Summary of the invention

[0005] In view of this, the present invention designs a training-free full-reference image quality evaluation method based on wavelet transform and Hausdorff distance, which extracts perceptual features of different frequency bands based on multi-scale wavelet transform, simulates the multi-scale processing characteristics of the human visual system (HVS), and adaptively adjusts low-frequency and high-frequency features through a collaborative feature optimization module to generate features that are more in line with perceptual correlation. The Hausdorff distance is used to quantify the difference in feature distribution between the reference image and the distorted image, avoiding the limitations of traditional point-to-point measurement.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A full-reference image quality evaluation method based on wavelet transform and Hausdorff distance comprises the following steps:

[0008] S1, obtaining a reference image and a distorted image, and performing normalization processing;

[0009] S2, decomposing the normalized reference image into a low-frequency sub-band image and a high-frequency sub-band image of the reference image through frequency domain transformation;

[0010] Decomposing the distorted image after normalization into a low-frequency sub-band image and a high-frequency sub-band image of the distorted image through frequency domain transformation;

[0011] S3, performing weighted optimization on the low-frequency sub-band image and the high-frequency sub-band image of the reference image to obtain the depth feature of the reference image;

[0012] The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are weighted optimized to obtain the depth feature of the distorted image;

[0013] S4, evaluating the distribution difference of the depth features of the test image and the depth features of the distorted image by Hausdorff distance, obtaining a Hausdorff distance that measures the difference information between the two sets of feature distributions, and calculating a feature similarity score according to the Hausdorff distance;

[0014] S5, calculating the distance between the color histograms of the reference image and the distorted image as the final similarity measurement weight;

[0015] S6. Fusion feature similarity scores and similarity measurement weights generate the final image quality score.

[0016] Furthermore, the step S2 specifically includes:

[0017] Perform wavelet transform on the normalized reference image to obtain the low-frequency subband image of the reference image and high frequency subband images

[0018] Perform wavelet transform on the distorted image after normalization to obtain the low-frequency sub-band image of the distorted image and high frequency subband images

[0019] Furthermore, in step S3, the low-frequency sub-band image and the high-frequency sub-band image of the reference image are weighted optimized to obtain the depth feature of the reference image, which is expressed as:

[0020]

[0021] Among them, F r represents the deep features of the reference image, and R represents the weighted optimization function;

[0022] The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are weighted optimized to obtain the deep features of the distorted image, which is expressed as:

[0023]

[0024] Among them, F r represents the deep features of the reference image, and R represents the weighted optimization function.

[0025] Furthermore, the step S4 uses the Hausdorff distance to evaluate the difference between the optimization features of the considered image and the optimization features of the distorted image, which specifically includes the following expression:

[0026]

[0027] Among them, H(F r ,F d ) represents the Hausdorff distance between the optimized features of the evaluation image and the optimized features of the distorted image, d γ (F r ,F d ) represents the joint distribution γ under F r ,F d The expected distance between feature distributions.

[0028] Furthermore, the step S6 of fusing the feature similarity score with the similarity measurement weight to generate the final image quality score includes constructing the following expression:

[0029]

[0030] Q=MAP g (sim Haus (F r ,F d ),X(H r ,H d ))

[0031] Among them, Avg(·) represents the mean of the difference in the distribution distances of the extracted 6 depth features of different scales, and H avg represents the distribution difference between the depth features of the reference image and the depth features of the distorted image, sim Haus (F r ,F d ) represents the similarity function used to measure the distribution of reference features and distorted features, MAP g Represents the function used to fuse feature similarity scores with similarity metric weights to produce the final quality score.

[0032] The present invention also discloses a full reference image quality evaluation system based on wavelet transform and Hausdorff distance, comprising:

[0033] A data preprocessing module is used to obtain the distorted image and the reference image and perform normalization processing;

[0034] A frequency domain transformation module, used for decomposing the normalized reference image into a low-frequency sub-band image and a high-frequency sub-band image of the reference image through frequency domain transformation;

[0035] Decomposing the distorted image after normalization into a low-frequency sub-band image and a high-frequency sub-band image of the distorted image through frequency domain transformation;

[0036] A collaborative feature refinement module is used to perform weighted optimization on the low-frequency sub-band image and the high-frequency sub-band image of the reference image to obtain the deep features of the reference image;

[0037] The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are weighted optimized to obtain the depth feature of the distorted image;

[0038] A feature similarity score calculation module is used to evaluate the difference between the optimized features of the test image and the optimized features of the distorted image using the Hausdorff distance, obtain the Hausdorff distance that measures the difference information between the two sets of feature distributions, and calculate the feature similarity score according to the Hausdorff distance;

[0039] The final similarity measurement weight acquisition module is used to calculate the distance C (H) between the color histogram of the reference image and the distorted image. r ,H d ) as the final similarity measurement weight;

[0040] The image quality score output module is used to fuse the feature similarity score and the similarity measurement weight and output the final image quality score.

[0041] Preferably, the frequency domain transformation module specifically performs the following steps:

[0042] Perform wavelet transform on the normalized reference image to obtain the low-frequency subband image of the reference image and high frequency subband images

[0043] Perform wavelet transform on the distorted image after normalization to obtain the low-frequency sub-band image of the distorted image and high frequency subband images

[0044] Preferably, in the collaborative feature refinement module, the following steps are specifically performed:

[0045] Decomposing the low-frequency sub-band image and the high-frequency sub-band image of the reference image respectively to obtain the corresponding detail coefficient components and approximate coefficient components;

[0046] Using a hyperbolic tangent function to enhance the perceptual relevance of each detail coefficient component, to obtain an enhanced detail coefficient component;

[0047] The adaptive weight of the enhanced detail coefficient component is calculated by using energy and logarithmic function, and the enhanced detail coefficient component is weighted to obtain the weighted detail coefficient component;

[0048] The weighted detail coefficient component and the approximate coefficient component are fused to obtain the weighted optimized depth feature of the reference image;

[0049] The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are decomposed to obtain the corresponding detail coefficient components and approximate coefficient components respectively;

[0050] Using a hyperbolic tangent function to enhance the perceptual relevance of each detail coefficient component, to obtain an enhanced detail coefficient component;

[0051] The adaptive weight of the enhanced detail coefficient component is calculated by using energy and logarithmic function, and the enhanced detail coefficient component is weighted to obtain the weighted detail coefficient component;

[0052] The weighted detail coefficient components are fused with the approximate coefficient components to obtain the weighted optimized depth features of the distorted image.

[0053] Preferably, the image quality score output module outputs the final image quality score, including constructing the following expression:

[0054]

[0055] Among them, Avg(·) represents the mean of the difference in the distribution distances of the extracted 6 depth features of different scales, and H avg represents the distribution difference between the depth features of the reference image and the depth features of the distorted image, sim Haus (F r ,F d ) represents the similarity function used to measure the distribution of reference features and distorted features, MAP g Represents the function used to fuse feature similarity scores with similarity metric weights to produce the final quality score.

[0056] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following beneficial effects:

[0057] The present invention provides a full reference image quality evaluation method based on wavelet transform and Hausdorff distance. By combining multi-scale wavelet transform and Hausdorff distance, the quality difference between the reference image and the distorted image can be accurately evaluated. The method first decomposes the reference image and the distorted image into multiple frequency sub-bands by wavelet transform, wherein the low-frequency sub-band mainly contains information related to color and brightness distortion, while the high-frequency sub-band captures the edge and texture distortion in the image. Next, the sub-band image after wavelet transform is further refined and decomposed into approximate coefficients C A , and detail coefficient C D , and further refine these two coefficients into four coefficient components CAA , C AD , C DA , C DD , in order to more accurately capture the distortion characteristics of the image in different frequency bands. Then, the Hausdorff distance is used to quantify the difference in feature distribution between the reference image and the distorted image, overcoming the limitations of traditional point-to-point measurement and being able to comprehensively evaluate the detail differences between images. In order to further enhance the perceptual relevance of the evaluation, this method calculates the distance C (H) between the color histograms of the reference image and the distorted image. r ,H d ) is used to provide the weight of the final similarity measure, ensuring that the evaluation results are more consistent with the human eye's perception of color distortion and detail distortion. In terms of feature optimization, the collaborative feature refinement module is used to adaptively weight low-frequency and high-frequency features, so that low-frequency features (such as color and brightness information) can be enhanced while suppressing irrelevant interference; high-frequency features (such as edge and texture information) are enhanced according to different distortion types. This process helps to improve the model's sensitivity to different types of distortion (such as compression, noise, etc.). Finally, by combining the outputs from the collaborative feature optimization module and the distribution similarity measurement module based on the Hausdorff distance, the final perceptual quality score Q is generated. This quality score accurately reflects the perceptual difference between the reference image and the distorted image, and can better simulate the human eye's intuitive perception of image quality. The entire method does not require training and has high robustness and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0059] Figure 1 A flow chart of a full-reference image quality evaluation method based on wavelet transform and Hausdorff distance provided by the present invention;

[0060] Figure 2 The architecture diagram of the full-reference image quality evaluation method based on wavelet transform and Hausdorff distance provided by the present invention; DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The embodiment of the present invention provides a full-reference image quality assessment method based on a local area mask, which can accurately assess image quality and is closer to image quality assessment perceived by the human eye. The specific embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0063] The embodiment of the present invention discloses a full reference image quality evaluation method based on wavelet transform and Hausdorff distance, such as Figure 1 As shown, the following steps are included:

[0064] S1, obtaining a reference image and a distorted image, and performing normalization processing;

[0065] S2, decomposing the normalized reference image into a low-frequency sub-band image and a high-frequency sub-band image of the reference image through frequency domain transformation;

[0066] Decomposing the distorted image after normalization into a low-frequency sub-band image and a high-frequency sub-band image of the distorted image through frequency domain transformation;

[0067] S3, performing weighted optimization on the low-frequency sub-band image and the high-frequency sub-band image of the reference image to obtain the depth feature of the reference image;

[0068] The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are weighted optimized to obtain the depth feature of the distorted image;

[0069] S4, evaluating the distribution difference of the depth features of the test image and the depth features of the distorted image by Hausdorff distance, obtaining a Hausdorff distance that measures the difference information between the two sets of feature distributions, and calculating a feature similarity score according to the Hausdorff distance;

[0070] S5, calculating the distance between the color histograms of the reference image and the distorted image as the final similarity measurement weight;

[0071] S6. Fusion feature similarity scores and similarity measurement weights generate the final image quality score.

[0072] The specific implementation steps of the present invention are described in detail below.

[0073] In step S1, first obtain the distorted image and the reference image, and use image processing software or library to uniformly resize the images to 256x256 pixels. This step ensures the consistency of subsequent processing and the effectiveness of comparison.

[0074] In step S2, the image is decomposed by frequency domain transformation. This embodiment adopts wavelet transformation method. Specifically, the normalized reference image and the distorted image are subjected to wavelet transformation respectively to extract multi-scale and multi-band image features and capture color, brightness, edge and texture information. The image is divided into low-frequency sub-bands W L and high frequency subband W H , where the low frequency subband W L Contains the color and brightness information of the image, the high-frequency subband W H Including the edge and texture features of the image, the above process can be expressed as:

[0075] W r =W(I r );

[0076] W d =W(I d );

[0077] Among them, I r is the reference image, I d is the distorted image, and W represents the wavelet transform operation.

[0078] In the above process, the low frequency subband W L Mainly contains low-low frequency component S LL ; and the high frequency subband W H Contains low-high frequency components S LH , high-low frequency component S HL , and high-high frequency component S HH , the above implementation process can be achieved through four convolution operations.

[0079] For the low-low frequency component S LL , including the approximation coefficient C AA ; while the low-high frequency component S LH , high-low frequency component S HL , and high-high frequency component S HH Contains detail coefficients C AD , C DA and C DD The above process can be expressed as:

[0080] [S LL ,S LH ,S HL ,S HH ]=Conv([f LL,f LH ,f HL ,f HH ],Img)

[0081] In the above formula, i and j represent the horizontal and vertical coordinates in the corresponding sub-band image respectively.

[0082]

[0083] Among them, S represents the distorted image or reference image, i, j represents the horizontal and vertical coordinate values ​​of the image pixel values, and the distorted image or reference image is decomposed into approximate coefficients C A and detail coefficient C D , C A and C D It is further decomposed into four coefficient components C AA , C AD , C DA , C DD .

[0084] In step S3, the four coefficient component features are weighted optimized to generate the deep features of the reference image. and deep features of distorted images The specific implementation method of the above process is as follows:

[0085] For the reference image or the distorted image, the hyperbolic tangent function is used to enhance the perceptual correlation of each detail coefficient component to obtain the enhanced detail coefficient component; the adaptive weight of the enhanced detail coefficient component is calculated using the energy and logarithmic function, and the enhanced detail coefficient component is weighted to obtain the weighted detail coefficient component; the weighted detail coefficient component is fused and reconstructed with the approximate coefficient component to obtain the weighted optimized depth feature.

[0086] In actual practice, deep features can be acquired with the help of the VGG-19 feature extraction network.

[0087] Further, in step S4, the refined reference image depth feature F r and distorted image depth feature F d The difference between the feature distributions of the reference image and the distorted image is evaluated by the Hausdorff distance. The formula is expressed as:

[0088]

[0089]

[0090] Among them, H(F r ,F d) represents the Hausdorff distance, which is used to measure the difference between the depth feature distribution of the reference image and the depth feature of the distorted image. Its upper bound can be expressed as d γ (F r ,F d ), d γ (F r ,F d ) represents the joint distribution γ under F r ,F d The expected distance between feature distributions.

[0091] In step S5, in order to further enhance the perceptual relevance of the similarity measure, the distance C (H r ,H d ) is used as the final similarity measurement weight, and the formula is expressed as:

[0092]

[0093] Among them, H r and H d Represent the histograms of the reference image and the distorted image respectively.

[0094] In step S6, the final perceptual quality score Q is estimated by combining the feature similarity score and the final similarity measurement weight, and a quality score that conforms to human perception is generated based on the feature distribution difference. It can be expressed as:

[0095]

[0096] Q=MAP g (sim Haus (F r ,F d ),X(H r ,H d ))

[0097] Among them, Avg(·) is used to find the mean of the difference in the distribution distances of the extracted 6 depth features of different scales, and H avg is the distribution difference between the reference feature and the distorted feature, sim Haus (F r ,F d ) is used to measure the similarity between the reference feature and the distorted feature distribution, MAP g Used to fuse feature similarity scores and similarity metric weights to produce the final quality score.

[0098] On the other hand, the present invention also discloses a full-reference image quality evaluation system based on wavelet transform and Hausdorff distance, comprising:

[0099] A data preprocessing module is used to obtain the distorted image and the reference image and perform normalization processing;

[0100] A frequency domain transformation module, used for decomposing the normalized reference image into a low-frequency sub-band image and a high-frequency sub-band image of the reference image through frequency domain transformation;

[0101] Decomposing the distorted image after normalization into a low-frequency sub-band image and a high-frequency sub-band image of the distorted image through frequency domain transformation;

[0102] A collaborative feature refinement module is used to perform weighted optimization on the low-frequency sub-band image and the high-frequency sub-band image of the reference image to obtain the deep features of the reference image;

[0103] The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are weighted optimized to obtain the depth feature of the distorted image;

[0104] A feature similarity score calculation module is used to evaluate the difference between the optimized features of the test image and the optimized features of the distorted image using the Hausdorff distance, obtain the Hausdorff distance that measures the difference information between the two sets of feature distributions, and calculate the feature similarity score according to the Hausdorff distance;

[0105] The final similarity measurement weight acquisition module is used to calculate the distance C (H) between the color histogram of the reference image and the distorted image. r ,H d ) as the final similarity measurement weight;

[0106] The image quality score output module is used to fuse the feature similarity score and the similarity measurement weight and output the final image quality score.

[0107] The overall architecture of the system is referenced Figure 2 ,The execution process of each module in the system is described in the ,reference method and will not be repeated here.

[0108] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A full-reference image quality assessment method based on wavelet transform and Hausdorff distance, characterized in that: The steps include: S1, obtaining a reference image and a distorted image, and performing normalization processing; S2, decomposing the normalized reference image into a low-frequency sub-band image and a high-frequency sub-band image of the reference image through frequency domain transformation; Decomposing the distorted image after normalization into a low-frequency sub-band image and a high-frequency sub-band image of the distorted image through frequency domain transformation; S3, performing weighted optimization on the low-frequency sub-band image and the high-frequency sub-band image of the reference image to obtain the depth feature of the reference image; The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are weighted optimized to obtain the depth feature of the distorted image; S4, evaluating the distribution difference of the depth features of the test image and the depth features of the distorted image by Hausdorff distance, obtaining a Hausdorff distance that measures the difference information between the two sets of feature distributions, and calculating a feature similarity score according to the Hausdorff distance; S5, calculating the distance between the color histograms of the reference image and the distorted image as the final similarity measurement weight; S6. Fusion feature similarity scores and similarity measurement weights generate the final image quality score.

2. A full-reference image quality assessment method based on wavelet transform and Hausdorff distance as claimed in claim 1, characterized in that: The step S2 specifically includes: Perform wavelet transform on the normalized reference image to obtain the low-frequency subband image of the reference image and high frequency subband images Perform wavelet transform on the distorted image after normalization to obtain the low-frequency sub-band image of the distorted image and high frequency subband images 3. A full reference image quality assessment method based on wavelet transform and Hausdorff distance as claimed in claim 2, characterized in that: In step S3, the low-frequency sub-band image and the high-frequency sub-band image of the reference image are weighted optimized to obtain the depth feature of the reference image, which is expressed as: Among them, F r represents the deep features of the reference image, and R represents the weighted optimization function; The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are weighted optimized to obtain the deep features of the distorted image, which is expressed as: Among them, F r represents the deep features of the reference image, and R represents the weighted optimization function.

4. A training-free full-reference image quality assessment method based on wavelet transform and Hausdorff distance as claimed in claim 1, characterized in that: The step S4 uses the Hausdorff distance to evaluate the difference between the optimization features of the considered image and the optimization features of the distorted image, which specifically includes the following expression: Among them, H(F r ,F d ) represents the Hausdorff distance between the optimized features of the evaluation image and the optimized features of the distorted image, d γ (F r ,F d ) represents the joint distribution γ under F r ,F d The expected distance between feature distributions.

5. The training-free full-reference image quality assessment method based on wavelet transform and Hausdorff distance as claimed in claim 1, characterized in that: The step S6 of fusing the feature similarity score with the similarity measurement weight to generate the final image quality score includes constructing the following expression: Q=MAP g (sim Haus (F r ,F d ),C(H r ,H d )) Among them, Avg(·) represents the mean of the difference in the distribution distances of the extracted 6 depth features of different scales, and H avg represents the distribution difference between the depth features of the reference image and the depth features of the distorted image, sim Haus (F r ,F d ) represents the similarity function used to measure the distribution of reference features and distorted features, MAP g Represents the function used to fuse feature similarity scores with similarity metric weights to produce the final quality score.

6. A full-reference image quality assessment system based on wavelet transform and Hausdorff distance, characterized in that: include: A data preprocessing module is used to obtain the distorted image and the reference image and perform normalization processing; A frequency domain transformation module, used for decomposing the normalized reference image into a low-frequency sub-band image and a high-frequency sub-band image of the reference image through frequency domain transformation; Decomposing the distorted image after normalization into a low-frequency sub-band image and a high-frequency sub-band image of the distorted image through frequency domain transformation; A collaborative feature refinement module is used to perform weighted optimization on the low-frequency sub-band image and the high-frequency sub-band image of the reference image to obtain the deep features of the reference image; The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are weighted optimized to obtain the depth feature of the distorted image; A feature similarity score calculation module is used to evaluate the difference between the optimized features of the test image and the optimized features of the distorted image using the Hausdorff distance, obtain the Hausdorff distance that measures the difference information between the two sets of feature distributions, and calculate the feature similarity score according to the Hausdorff distance; The final similarity measurement weight acquisition module is used to calculate the distance C (H) between the color histogram of the reference image and the distorted image. r ,H d ) as the final similarity measurement weight; The image quality score output module is used to fuse the feature similarity score and the similarity measurement weight and output the final image quality score.

7. A full reference image quality assessment system based on wavelet transform and Hausdorff distance according to claim 6, characterized in that: The frequency domain transformation module specifically performs the following steps: Perform wavelet transform on the normalized reference image to obtain the low-frequency subband image of the reference image and high frequency subband images Perform wavelet transform on the distorted image after normalization to obtain the low-frequency sub-band image of the distorted image and high frequency subband images 8. A full reference image quality assessment system based on wavelet transform and Hausdorff distance according to claim 6, characterized in that: In the collaborative feature refinement module, the following steps are specifically performed: Decomposing the low-frequency sub-band image and the high-frequency sub-band image of the reference image respectively to obtain the corresponding detail coefficient components and approximate coefficient components; Using a hyperbolic tangent function to enhance the perceptual relevance of each detail coefficient component, to obtain an enhanced detail coefficient component; The adaptive weight of the enhanced detail coefficient component is calculated by using energy and logarithmic function, and the enhanced detail coefficient component is weighted to obtain the weighted detail coefficient component; The weighted detail coefficient component and the approximate coefficient component are fused to obtain the weighted optimized depth feature of the reference image; The low-frequency sub-band image and the high-frequency sub-band image of the distorted image are decomposed to obtain the corresponding detail coefficient components and approximate coefficient components respectively; Using a hyperbolic tangent function to enhance the perceptual relevance of each detail coefficient component, to obtain an enhanced detail coefficient component; The adaptive weight of the enhanced detail coefficient component is calculated by using energy and logarithmic function, and the enhanced detail coefficient component is weighted to obtain the weighted detail coefficient component; The weighted detail coefficient components are fused with the approximate coefficient components to obtain the weighted optimized depth features of the distorted image.

9. A full reference image quality assessment system based on wavelet transform and Hausdorff distance according to claim 6, characterized in that: The image quality score output module outputs the final image quality score, including constructing the following expression: Q=MAP g (sim Haus (F r ,F d ),C(H r ,H d )) Among them, Avg(·) represents the mean of the difference in the distribution distances of the extracted 6 depth features of different scales, and H avg represents the distribution difference between the depth features of the reference image and the depth features of the distorted image, sim Haus (F r ,F d ) represents the similarity function used to measure the distribution of reference features and distorted features, MAP g Represents the function used to fuse feature similarity scores with similarity metric weights to produce the final quality score.

Images