No-reference vehicle image quality evaluation method and equipment based on saliency perception

Abstract

The invention discloses a no-reference vehicle image quality evaluation method and equipment based on saliency perception. A large-scale no-label vehicle image data set is utilized, and image distortion recognition models for distinguishing different types and grades are trained in a self-supervision mode; training a model for detecting the salient region in the image by using the salient region segmentation data set; a pre-trained distortion quality perception module and a pre-trained saliency region segmentation module are used to acquire a saliency feature map and a distortion feature map in an input image, and then a weight matrix is constructed and applied to realize weighted fusion between the saliency feature map and the distortion feature map so as to generate a new comprehensive feature map; and sending the generated comprehensive feature map into a multi-layer perceptron, and outputting a score reflecting the overall quality of the image after processing. The method not only improves the accuracy of image quality evaluation, but also enhances the practical application value and generalization ability of the model. The method is of great significance in key tasks such as pedestrian detection and traffic sign recognition.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a reference-free vehicle image quality assessment method and device based on saliency perception. Background Art

[0002] Deep learning belongs to the category of artificial intelligence and machine learning. It solves real-world problems by building a neural network structure, performing feature extraction and regression decision-making. In recent years, deep learning has made significant progress in the field of image quality assessment, especially in the automotive industry. Models based on convolutional neural networks (CNNs) can automatically evaluate the quality of images after training on large-scale subjective databases. These models can largely simulate human perception of image quality by learning features from a large amount of image data, which is particularly important for autonomous driving systems. For example, in an automotive environment, images captured by on-board cameras are used to support various advanced driver assistance systems (ADAS), such as lane keeping assist, adaptive cruise control, and automatic emergency braking. High-quality images are essential to ensure the accuracy and reliability of these systems. However, existing image quality assessment methods based on deep learning, such as CN118608927A, have the following problems in automotive application scenarios: 1. Existing methods usually process the entire image indiscriminately and cannot distinguish different areas in the image. This processing method is inconsistent with the processing mechanism of the human visual system, and the human eye will give priority to salient areas in the scene. In the automotive field, this may mean that the system fails to correctly identify road signs, pedestrians, or other important obstacles because these key elements are often located in specific parts of the image rather than evenly distributed; 2. Because the differences in the importance of various regions in the image are not taken into account, existing methods may not accurately reflect the overall quality of the image. For example, the blur of foreground areas (such as pedestrians or obstacles in front of the vehicle) usually has a greater impact on image quality than the blur of background areas (such as distant scenery). For autonomous driving systems, this defect may lead to potential safety hazards because it may underestimate the quality loss of important visual information. In order to meet the above challenges, new image quality evaluation methods need to understand and process different areas of the image more intelligently, imitate the human visual system's attention to salient areas, and evaluate image quality weightedly according to the actual importance of each area. Such improvements are of great significance to improving the safety and reliability of autonomous driving technology. Summary of the invention

[0003] The present invention proposes a method, device and storage medium for non-reference vehicle image quality assessment based on saliency perception, which can solve at least one of the technical problems in the background technology; the method is suitable for image quality evaluation in various practical application scenarios.

[0004] To achieve the above object, the present invention adopts the following technical solutions: A no-reference vehicle image quality assessment method based on saliency perception comprises the following steps: S1, uses a large-scale unlabeled vehicle image dataset and trains a self-supervised image distortion recognition model to distinguish different types and levels; S2, uses the salient region segmentation dataset to train a model dedicated to detecting salient regions in an image, and in subsequent stages focuses on those parts that are more important to the human visual system; S3, for the input image, uses the distortion quality perception module pre-trained in S1 and the salient region segmentation module pre-trained in S2 to obtain the salient feature map and the distortion feature map in the input image, and then realizes the weighted fusion between the two by constructing and applying an appropriate weight matrix to generate a new comprehensive feature map; S4, sends the comprehensive feature map generated by S3 to a multi-layer perceptron (MLP), which outputs a score reflecting the overall quality of the image after further processing.

[0005] Furthermore, in step S1, specifically: S11, collects a large-scale unlabeled vehicle image dataset, including synthetic distorted images and real distorted images. The dataset covers various types of distortions, such as noise, blur, and compression artifacts; S12, design an image distortion module based on convolutional neural network (CNN). This module uses ResNet-50 as the backbone network and is followed by a multi-layer perceptron (MLP) head to extract high-level features of the image; S13, train the image distortion module using a self-supervised learning method. Specifically, the model is trained by contrastive learning so that the model can recognize and adapt to various categories and levels of image quality degradation, and a self-supervised loss function is defined. The InfoNCE loss function is used to calculate the loss between the representations generated by the query encoder and the key encoder. The InfoNCE loss function can be expressed as: .

[0006] Furthermore, in step S2, specifically: S21, collects a salient region segmentation dataset containing a large number of natural vehicle images and their corresponding salient region annotations; S22, build a hierarchical salient feature perception module based on the U-Net architecture, which mainly consists of encoder and decoder components. The encoder stage includes residual U blocks (RSU) of different depths, such as RSU-7, RSU-6, RSU-5, RSU-4, and the expanded version of RSU-4F, to adapt to input feature maps of different resolutions. The decoder stage adopts corresponding upsampling operations, and at some stages, dilated convolutions are introduced to prevent the loss of context information during downsampling; S23, uses the input convolution layer to convert the original image into an intermediate feature map to facilitate local feature extraction. The multi-scale context information is extracted from the intermediate feature map through a U-Net structure with a height L, and the local features and multi-scale features are merged through residual connections to generate a comprehensive salient feature representation; S24, train the model on the prepared salient region segmentation dataset, and the optimization goal is to enable the model to accurately segment the salient regions in the image. Use cross entropy loss as the optimization criterion.

[0007] The Adam optimization algorithm is used to update the parameters. After training, the model parameters are frozen so that they can be integrated into the subsequent image quality assessment process. Furthermore, in step S3, a significant feature map and a distortion feature map are extracted from the input image respectively, and a weighted fusion between the two is realized by constructing and applying an appropriate weight matrix to generate a new comprehensive feature map. S And distortion feature map Q , and construct a trainable weight matrix W Perform weighted processing on the salient feature map.

[0008] Furthermore, in step S4, specifically: S41, the fused comprehensive feature map is sent to a multi-layer perceptron head to further refine and integrate the features. The final image quality score is output through the MLP head. The MLP head is usually composed of multiple fully connected layers, and each fully connected layer except the last layer is followed by a nonlinear activation function. The specific formula for the final quality score q is as follows:

[0009] in, and They are The weight matrix and bias vector of the layer, is the number of layers of the MLP.

[0010] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0011] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0012] It can be seen from the above technical solution that the saliency perception-based no-reference vehicle image quality assessment method of the present invention is specifically a saliency perception no-reference image quality assessment method integrating distortion category information. The method aims to more accurately assess image quality and make it more consistent with human subjective judgment by introducing a hierarchical saliency perception module and a self-supervised learning quality perception module.

[0013] The present invention provides a saliency-aware reference-free image quality assessment method that integrates distortion category information. The method aims to more accurately assess image quality and make it more consistent with human subjective judgment by introducing a hierarchical saliency perception module and a quality perception module of self-supervised learning. Specifically, the method of the present invention can effectively distinguish different areas in an image, emphasize the importance of salient areas, and reduce the influence of background areas. This method not only improves the accuracy of image quality assessment, but also enhances the generalization ability of the model in practical applications, especially in the absence of labeled data. In this way, the working mechanism of the human visual system is simulated, thereby achieving image quality assessment that is closer to human perception.

[0014] Specifically, the present invention utilizes a hierarchical salient feature perception module based on the U-Net architecture, which can effectively distinguish different areas in the image, especially emphasize the importance of salient areas. This module ensures high-precision recognition of visually important parts of the image through multi-scale context information extraction and residual connection mechanism, reducing the impact of background areas on the evaluation results. In addition, by designing a model specifically for detecting salient areas in images, the present invention simulates the attention mechanism of the human visual system. This design allows the evaluation process to focus not only on the overall distortion of the image, but also pays special attention to those parts that are critical to human perception quality, achieving an image quality evaluation that is closer to human perception.

[0015] In order to further improve the generalization ability and adaptability, the present invention introduces a self-supervised learning method to train the image distortion recognition module. This enables the present invention to maintain efficient performance in the absence of labeled data. Self-supervised learning uses the InfoNCE loss function through a contrastive learning framework to enhance the adaptability and robustness of the model to various types and degrees of distortion, greatly improving the generalization ability of the model in different application scenarios. Combined with distortion category information and saliency perception, the present invention can provide a more detailed and comprehensive image quality assessment. By extracting saliency feature maps and distortion feature maps from the input image respectively, and performing weighted fusion to generate a comprehensive feature map, the multi-layer perceptron (MLP) finally outputs a score reflecting the overall quality of the image, ensuring the accuracy and reliability of the evaluation results.

[0016] It is worth mentioning that the application of the self-supervised learning mechanism significantly reduces the demand for large-scale annotated data sets, saving a lot of manpower and time costs, allowing the present invention to be effectively trained and evaluated on more diverse image data, further expanding its scope of application. In summary, the saliency-aware reference-free image quality assessment method that integrates distortion category information proposed in the present invention not only improves the accuracy of image quality assessment, but also enhances the practical application value and generalization ability of the model. In this way, the present invention provides a more intelligent, efficient and human visual habit-compliant quality assessment solution for the field of image processing.

[0017] In the automotive industry, high-quality images are essential to ensure the accuracy and reliability of advanced driver assistance systems (ADAS) and autonomous driving systems. Images captured by on-board cameras are not only used for navigation and environmental perception, but also for key tasks such as pedestrian detection and traffic sign recognition. Therefore, the point image quality assessment of the present invention is extremely important in this field. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is the overall framework diagram of the method of the present invention; Figure 2 is a flow chart of a pre-trained image distortion recognition model in an embodiment of the present invention; Figure 3 An architecture diagram of a pre-trained saliency region perception module in an embodiment of the present invention; Figure 4 The evaluation results of the vehicle reference image, background blurred image and subject blurred image of this embodiment are respectively shown. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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.

[0020] This embodiment will be described based on the LIVE dataset. The LIVE dataset was created by the Image and Video Engineering Laboratory of the University of Texas at Austin in 2006. The reference images of the entire dataset are 29 high-resolution and high-quality color images collected from the Internet and photography CDs, including face images, animal images, close-up images, wide-angle images, natural scenes, man-made objects, and images with different foreground / background configurations.

[0021] The method of the present invention first uses a large-scale image dataset and adopts the MOCO (Momentum Contrast) architecture to train an image distortion perception module. Through self-supervised learning, the present invention can effectively deal with the problem of scarce labeled data. Next, in order to capture the content differences between different salient regions, the present invention further develops an image saliency perception module. The features extracted by the two modules are then weighted and integrated to obtain the final fusion features. In this process, the present invention selects a multi-layer perceptron (MLP) as the last part of the model, which is responsible for outputting the quality score based on the above-mentioned fusion features. With its powerful nonlinear fitting ability, MLP can accurately map the complex relationship from input features to image quality scores, ensuring the accuracy and reliability of the evaluation results.

[0022] The following are specific instructions: The embodiment of the present invention provides a no-reference vehicle image quality assessment method based on saliency perception, such as Figure 1 As shown, the specific steps are as follows: S1,pretrains the quality-aware model through self-supervised learning method using large-scale unlabeled datasets (such as KADID, AVA, COCO, etc.).,This model uses ResNet-50 as the backbone network and is followed by a multi-layer perceptron (MLP) head.

[0023] S2, using the DUTS-TR dataset (including 21,106 images augmented by horizontal flipping), trains a hierarchical salient feature perception module based on the U-Net architecture. This model is specifically designed to detect salient regions in images.

[0024] S31, for the input image, first use the quality-aware model pre-trained in S1 to extract quality features. Assume that the model output is , and the present invention selects the output of the last mapping head and flattens it into a one-dimensional vector with a shape of [2048, 1]. This step can be expressed as:

[0025] in, is the input image, represents the quality perception model, Represents the operation of flattening a multidimensional output into a one-dimensional vector.

[0026] S32 For the input image, use the hierarchical salient feature perception module based on the U-Net architecture and trained on the DUTS-TR dataset in step S2 to extract salient features.

[0027] S33, weighted fusion is performed on the two feature vectors obtained in S32 and S31. First, the two feature vectors are concatenated together to form a new feature vector, and then weighted processing is performed through a fully connected layer (MLP). The whole process can be expressed by the following formula:

[0028] in, Indicates that and The result of splicing along the channel dimension is, It is a multi-layer perceptron, which is responsible for further processing and weighting the concatenated feature vectors.

[0029] S41, feature vector obtained by MLP It is sent to the last fully connected layer and outputs a score q that reflects the overall quality of the image. This process can be expressed by the following formula:

[0030] in, and are the weight matrix and bias vector of the last layer of MLP, is the output of the previous layer.

[0031] Dataset Introduction The LIVE dataset includes 779 distorted images. The dataset uses five computer distortion operations to degrade the reference images by 5 to 6 levels, resulting in 779 distorted images.

[0032] The TID2013 dataset contains more than 3,000 test images obtained from 25 reference images, each of which has 24 types of distortion, and each type of distortion is divided into 5 levels.

[0033] The CSIQ dataset contains 30 original images and 866 synthetically distorted images with six distortion types.

[0034] The KonIQ-10 dataset selects 10,073 images from the YFCC100M dataset. For each image, 120 reliable quality ratings were obtained through crowdsourcing, annotated by 1,459 volunteers, and the collected subjective data is as high as 1.2 million.

[0035] The LIVEC dataset contains 1,162 images captured with a wide variety of real distortions using modern mobile devices. These photos come from different photographers using different camera equipment in the real world, so these images contain complex real distortions.

[0036] The KADID-10K dataset contains 10,125 distorted images derived from 81 reference images. The dataset contains 25 different types of image distortions, including Gaussian blur, JPEG compression, added noise, chroma distortion, etc. Each distortion type is applied to different images to varying degrees.

[0037] Experimental analysis The model is validated on multiple widely used image quality assessment (IQA) datasets, including synthetic datasets (such as LIVE IQA, TID2013, CSIQ) and real distortion datasets (such as KonIQ-10K, CLIVE, SPAQ). The performance of the model is evaluated by indicators such as Pearson linear correlation coefficient (PLCC) and Spearman rank correlation coefficient (SRCC). The calculation formulas for PLCC and SRCC are:

[0038]

[0039] in, and are the predicted score and the true score respectively. and are their average values, The ranking difference between the measured score and the true score.

[0040] In order to demonstrate the effectiveness of the proposed algorithm, the evaluation indicators of several other algorithms are compared as shown in Table 1, and the best results are indicated in bold.

[0041] Table 1. Comparison of SRCC and PLCC of different algorithms

[0042] In summary, the present invention is closer to the human visual system by learning salient region features and distortion-like features separately. By learning salient region-specific and distortion-specific features in a separate manner, the hierarchical approach of the present invention is better able to capture quality degradation localized in salient image regions. As demonstrated by the results shown in Table 1, among these methods, the best performing method is marked in bold, while the second and third best performing methods are underlined. It is worth noting that the methods of the present invention achieved the third, first, and third positions on the synthetic dataset, respectively. On the real dataset, they all achieved the first place. This observation emphasizes the effectiveness of combining distortion-specific feature extraction with salient region feature extraction for image quality assessment across a variety of datasets.

[0043] In the automotive industry, high-quality images are essential to ensure the accuracy and reliability of advanced driver assistance systems (ADAS) and autonomous driving systems. Images captured by on-board cameras are not only used for navigation and environmental perception, but also for key tasks such as pedestrian detection and traffic sign recognition. Therefore, image quality assessment is of great significance in this field. In order to verify the effectiveness of the method of the present invention in actual application scenarios, the embodiments of the present invention are tested on multiple automotive-related datasets 1. Kitti dataset: a dataset widely used in autonomous driving research, containing a large number of images and videos of real-world driving scenes, covering different weather conditions and lighting environments. 2. Cityscapes dataset: focusing on images of urban street scenes, providing detailed semantic segmentation annotations, suitable for evaluating the model's ability to recognize salient areas in complex urban environments. The method of the present invention not only performs well on synthetic datasets, but also demonstrates excellent performance in real driving scenarios. In particular, on the Kitti and Cityscapes datasets, the method of the present invention can accurately capture the quality degradation of salient areas such as road signs, pedestrians and other obstacles, which is particularly important in autonomous driving systems.

[0044] like Figure 4As shown, the evaluation results of the vehicle reference image, the background blurred image and the subject blurred image are shown respectively. The method of the present invention gives a lower quality score to the image with blurred vehicle subject by finely capturing the difference of blur in different regions, so as to be more in line with the actual needs and the perceptual characteristics of the human visual system. Specifically, in the reference image, the other methods scored 70.37, and the method of the present invention reached a higher score of 71.88, which is close to the human subjective score of 71.43. This shows that the evaluation of high-quality images by the present invention is more accurate and can better reflect the real image quality. In the background blurred image, other methods gave a higher score of 66.37, but the method of the present invention scored 64.33, which is more consistent with the human subjective score of 64.4. This result avoids the overestimation of the background blurred image and ensures the accuracy of the evaluation results. In the subject blurred image, the method of the present invention scored 56.50, which is closer to the human subjective score of 54.8, while other methods gave an excessively high score of 60.72. This shows the advantages of the present invention in evaluating severely distorted images, especially for the quality degradation of key areas (such as the vehicle body), providing a more reliable evaluation.

[0045] In summary, when faced with a reference image and an image with a blurred vehicle body and a blurred background, the method of the present invention can more finely capture the difference in blur in different regions, thereby giving a lower quality score to the image with a blurred vehicle body. This method not only better meets actual needs, but also better simulates the perceptual characteristics of the human visual system.

[0046] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0047] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0048] In another embodiment provided by the present application, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute any of the saliency perception-based no-reference vehicle image quality assessment methods in the above embodiments.

[0049] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above methods.

[0050] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0051] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0052] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0053] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A no-reference vehicle image quality assessment method based on saliency perception, characterized in that: The following steps are included: S1, uses a large-scale unlabeled vehicle image dataset and trains a self-supervised image distortion recognition model to distinguish different types and levels; S2, uses the salient region segmentation dataset to train a model for detecting salient regions in images; S3, input image, use the distortion quality perception module pre-trained in S1 and the salient region segmentation module pre-trained in S2 to obtain the salient feature map and the distortion feature map in the input image, and then achieve weighted fusion between the two by constructing and applying a weight matrix to generate a new comprehensive feature map; S4, sends the comprehensive feature map generated by S3 to a multi-layer perceptron, which outputs a score reflecting the overall quality of the image after further processing.

2. The method for non-reference vehicle image quality assessment based on saliency perception according to claim 1, characterized in that: In the step S1, specifically: S11, collects a large-scale unlabeled vehicle image dataset, including synthetic distorted images and real distorted images; S12, design an image distortion module based on convolutional neural network; the image distortion module shown uses ResNet-50 as the backbone network and is followed by a multi-layer perceptron head to extract high-level features of the image; S13, use the self-supervised learning method to train the image distortion module; specifically, the model is trained by contrastive learning, so that the model can recognize and adapt to various categories and levels of image quality degradation, define a self-supervised loss function, use the InfoNCE loss function, calculate the loss between the representations generated by the query encoder and the key encoder, for the positive sample pair and negative samples , the InfoNCE loss function is expressed as: ; in, represents the cosine similarity, is the temperature hyperparameter.

3. The saliency-aware no-reference vehicle image quality assessment method according to claim 2, characterized in that: In the step S2, specifically: S21, collecting a salient region segmentation dataset including natural vehicle images and their corresponding salient region annotations; S22, constructing a hierarchical salient feature perception module based on a U-Net architecture, wherein the hierarchical salient feature perception module includes an encoder and a decoder component; the encoder stage includes residual U blocks of different depths to adapt to input feature maps of different resolutions; The decoder stage adopts the corresponding upsampling operation, while the dilated convolution is introduced in the setup stage to prevent the loss of context information during the downsampling process; S23, uses the input convolution layer to convert the original image into an intermediate feature map to facilitate local feature extraction; extracts multi-scale context information from the intermediate feature map through a U-Net structure with a height L, and merges the local features and multi-scale features through residual connections to generate a comprehensive salient feature representation; S24, train the model on the prepared salient region segmentation dataset. The optimization goal is to enable the model to accurately segment the salient regions in the image. Use cross entropy loss as the optimization criterion. is the true saliency label, is the predicted significance probability distribution, then the cross entropy loss can be expressed as: And use Adam optimization algorithm to update parameters; After training is complete, the model parameters are frozen to facilitate its integration into the subsequent image quality assessment process.

4. The method for non-reference vehicle image quality assessment based on saliency perception according to claim 3 is characterized in that: In step S3, a significant feature map and a distortion feature map are extracted from the input image respectively, and a weighted fusion between the two is realized by constructing and applying an appropriate weight matrix to generate a new comprehensive feature map. S And distortion feature map Q , and construct a trainable weight matrix W Perform weighted processing on the salient feature map.

5. The method for non-reference vehicle image quality assessment based on saliency perception according to claim 4 is characterized in that: The step S4 specifically includes: The fused comprehensive feature map is sent to a multi-layer perceptron head to further refine and integrate the features; the final image quality score is output through the multi-layer perceptron head; the multi-layer perceptron head consists of multiple fully connected layers, and each fully connected layer except the last layer is followed by a nonlinear activation function; the specific formula for the final quality score q is as follows: in, and They are The weight matrix and bias vector of the layer, is the number of layers in the multilayer perceptron.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 5.

Images