Image quality evaluation method, system, device and medium for AI image

Through a multi-scale image quality evaluation model, combined with feature extraction and adaptive feature fusion modules, the content consistency, visual quality and authenticity of AI images are evaluated, which solves the shortcomings of AI image quality evaluation in the existing technology and realizes AI generation Multi-dimensional evaluation and efficient scoring of images.

CN118154571BActive Publication Date: 2024-11-01SHENZHEN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410385097.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-01
Publication Date
2024-11-01
Estimated Expiration
2044-04-01

AI Technical Summary

Technical Problem

The prior art is difficult to effectively evaluate the quality of AI-generated images, especially in terms of poor performance in content consistency, visual quality and authenticity, and traditional methods cannot be applied to individual images, and lack in-depth quality evaluation methods for AI images .

Method used

Using a multi-scale image quality evaluation model, the feature extraction module, adaptive feature fusion module and evaluation module are used to combine AI images and text prompt data of different scales to evaluate the content consistency, visual quality and authenticity of AI images, and use pre-processing. The technology generates multi-scale features and performs mass score predictions.

Benefits of technology

The multi-dimensional quality evaluation of AI-generated images is realized, showing strong performance and generalization capabilities, maintaining good robustness on different data sets, and able to effectively measure the semantic differences between text prompts and images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118154571B_ABST
    Figure CN118154571B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of image quality evaluation technology, and discloses an image quality evaluation method, system, device and medium for AI images, including: obtaining an initial AI image and corresponding text prompt data; preprocessing the initial AI image to obtain several AI images of different scales; inputting the AI ​​images of different scales and the corresponding text prompt data into a multi-scale image quality evaluation model for quality evaluation to obtain an image quality evaluation score; wherein the image quality evaluation score includes a content consistency score, a visual quality score and a authenticity score of the AI ​​image; the multi-scale image quality evaluation model includes a feature extraction module, an AFF module and an evaluation module connected in sequence; the feature extraction module is constructed based on an image encoder, and the evaluation module includes several fully connected layers arranged in parallel. The technical solution of the present invention shows strong performance in the multi-dimensional evaluation of AI-generated images.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of image quality evaluation, and in particular relates to an image quality evaluation method, system, device and medium for AI images. Background Art

[0002] In recent years, artificial intelligence generated content (AIGC) is quietly leading a profound change, which is reshaping and even subverting the production and consumption mode of digital content. As the main branch of AIGC, artificial intelligence generated images (AGI) show great application potential in all aspects of human life. The creation of AGI involves inputting text prompts into text-generated graph models, or inputting images into graph-generated image models. However, due to technical limitations, there are large quality differences between different AGIs, which need to be manually selected before use, which greatly limits the development of generation technology.

[0003] Unlike natural scene images (NSI) captured by cameras, AGI is generated directly by AI generative models. In general, NSI is usually affected by various distortions (e.g., compression, blur, noise, etc.), and the perceptual quality is scored based on the level of visual experience. In contrast, the quality definition and expression of AGI is different, and it is usually evaluated in a multi-dimensional way, mainly including visual experience, authenticity, and content consistency. For AGI, the scores of these three dimensions may also be different. For example, for some generated images, there is no obvious distortion, and the image content is consistent with the text description. However, the image does not show enough details, making it easy to identify as fake. Therefore, the visual experience and consistency of such images are scored higher, while the authenticity is scored lower. In practical applications, in addition to the visual experience affected by distortion and the authenticity affected by artistic expression, users are also concerned about the matching degree of the generated AGI in the task, that is, content consistency. Since traditional NSI-oriented IQA methods can only evaluate image quality in the dimension of visual experience, they are not suitable for the quality assessment task of AGI.

[0004] Research on AGI quality assessment is still in its infancy and has made very limited progress. A widely used strategy for evaluating AGI quality is to calculate the distance between the NSI group and the AGI group. However, this strategy must evaluate a set of images and is not applicable to the case of only one image. Therefore, more advanced methods should be proposed. Following the general steps of image quality assessment tasks, some quality assessment databases are first created based on subjective experiments to promote the development of objective image quality assessment methods. These databases should contain text cues and multi-dimensional quality labels. In addition to creating these databases, some mainstream NSI-oriented image quality assessment methods need to be tested, that is, they are retrained multiple times using quality labels of specific dimensions. The results show that most existing methods perform poorly, especially in the task of evaluating content consistency, because they only take images as input and cannot measure the degree of mismatch between text cues and generated images. Recently, some studies have proposed using different encoders to process text cues and images separately, and concatenate the features extracted by the two encoders to generate quality scores. However, directly concatenating features cannot effectively measure the semantic difference between text cues and images. In order to measure the alignment between text prompts and images in detail, some studies have tried to segment text prompts into multiple morphogens, split images into multiple patches, and calculate the alignment scores between morphogens and sub-images one by one. However, image segmentation is highly dependent on the designer's experience, and it is not clear how to establish the correspondence between morphogens and sub-images. In general, current research mainly stays at using some mainstream image quality assessment methods to benchmark the AGI database, and lacks in-depth research on designing AGI quality assessment methods. Summary of the invention

[0005] The purpose of the present invention is to provide an image quality evaluation method, system, device and medium for AI images to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides an image quality evaluation method for an AI image, comprising:

[0007] Get the initial AI image and corresponding text prompt data;

[0008] Preprocessing the initial AI image to obtain a plurality of AI images of different scales;

[0009] AI images of different scales and corresponding text prompt data are input into a multi-scale image quality assessment model for quality assessment to obtain an image quality assessment score; wherein the image quality assessment score includes a content consistency score, a visual quality score and an authenticity score of the AI ​​image; the multi-scale image quality assessment model includes a feature extraction module, an AFF module and an evaluation module connected in sequence; the feature extraction module is constructed based on an image encoder, and the evaluation module includes a number of fully connected layers arranged in parallel.

[0010] Optionally, the labels of the initial AI images include a visual quality score, a authenticity score, and a consistency score.

[0011] Optionally, preprocessing the initial AI image to obtain a preprocessed initial AI image specifically includes:

[0012] The initial AI image is scaled to obtain a plurality of AI images of different scales corresponding to the initial AI image.

[0013] Optionally, the training process of the multi-scale image quality assessment model includes:

[0014] Acquire training data; the training data includes AI training images and corresponding image quality evaluation scores;

[0015] An initial multi-scale image quality assessment model is constructed, the training data is input into the initial multi-scale image quality assessment model for quality assessment, and training is performed with the goal of minimizing the loss between the initial training result after quality assessment and the AI ​​training image and the corresponding image quality assessment score to obtain the multi-scale image quality assessment model.

[0016] Optionally, the processing process of the multi-scale image quality assessment model includes:

[0017] Inputting AI images of different scales and corresponding text prompt data into the feature extraction module to perform multi-scale feature extraction, thereby obtaining multi-scale semantic feature data and text feature data;

[0018] Inputting the multi-scale semantic feature data into the AFF module for adaptive feature fusion to obtain fused picture feature data;

[0019] Inputting the fused image feature data into the evaluation module for score prediction to obtain a visual quality score and authenticity score corresponding to the initial AI image;

[0020] The content consistency score corresponding to the initial AI image is calculated based on the fused image feature data and the text feature data.

[0021] An image quality evaluation system for AI images, comprising:

[0022] A data acquisition module is used to obtain an initial AI image and corresponding text prompt data; pre-process the initial AI image to obtain a number of AI images of different scales;

[0023] An image quality assessment module is used to input AI images of different scales and corresponding text prompt data into a multi-scale image quality assessment model for quality assessment to obtain an image quality assessment score; wherein the image quality assessment score includes a content consistency score, a visual quality score and an authenticity score of the AI ​​image; the multi-scale image quality assessment model includes a feature extraction module, an AFF module and an evaluation module connected in sequence; the feature extraction module is constructed based on an image encoder, and the evaluation module includes a number of fully connected layers arranged in parallel.

[0024] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program so that the electronic device executes the image quality assessment method of an AI image.

[0025] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the image quality evaluation method of an AI image is implemented.

[0026] The technical effects of the present invention are:

[0027] The present invention takes into account the characteristics of AI-generated images, and constructs an evaluation model based on its multi-dimensional evaluation characteristics. It inputs AI-generated images of different scales and corresponding text prompts to obtain quality scores in three dimensions: content consistency, visual quality, and authenticity. The present method shows strong performance in the multi-dimensional evaluation of AI-generated images. The present invention has good generalization ability on different data sets, and still maintains good robustness when tested on different data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] 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 will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0030] Figure 1Generate an image evaluation network structure diagram for artificial intelligence in an embodiment of the present invention;

[0031] Figure 2 is a structural diagram of an adaptive feature fusion (AFF) module in an embodiment of the present invention;

[0032] Figure 3 4 is a flowchart of image quality evaluation in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0034] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that each intermediate value between the upper and lower limits of the scope is also specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0035] Unless otherwise indicated, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art generally understood by the present invention. Although the present invention describes only preferred methods, any method similar or equivalent to that described herein may also be used in the implementation or testing of the present invention. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods related to the documents. In the event of a conflict with any incorporated document, the content of this specification shall prevail.

[0036] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present invention description without departing from the scope or spirit of the present invention. Other embodiments derived from the present invention description will be apparent to those skilled in the art. The present application description and examples are exemplary only.

[0037] The words “include,” “including,” “have,” “contain,” etc. used in this article are open-ended terms, meaning including but not limited to.

[0038] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0039] Embodiment 1

[0040] like Figure 1 - Figure 3 As shown, this embodiment provides an image quality evaluation method for an AI image, including: obtaining an initial AI image and corresponding text prompt data;

[0041] Preprocessing the initial AI image to obtain a plurality of AI images of different scales;

[0042] AI images of different scales and corresponding text prompt data are input into a multi-scale image quality assessment model for quality assessment to obtain an image quality assessment score; wherein the image quality assessment score includes a content consistency score, a visual quality score and an authenticity score of the AI ​​image; the multi-scale image quality assessment model includes a feature extraction module, an AFF module and an evaluation module connected in sequence; the feature extraction module is constructed based on an image encoder, and the evaluation module includes a number of fully connected layers arranged in parallel.

[0043] This embodiment takes into account the characteristics of AI-generated images, and constructs an evaluation model based on its multi-dimensional evaluation characteristics. It inputs AI-generated images of different scales and corresponding text prompts to obtain quality scores in three dimensions: content consistency, visual quality, and authenticity. This method shows strong performance in the multi-dimensional evaluation of AI-generated images. The present invention has good generalization ability on different data sets, and still maintains good robustness when tested on different data sets.

[0044] To advance the development of this field, this paper proposes a novel Adaptive Mixed-Scale Feature Fusion Network (AMFF-Net) for AGI quality assessment. Specifically, AMFF-Net adopts a multi-task framework to evaluate the quality of AGI from three dimensions: distortion, authenticity, and content consistency. Considering that the subjective evaluation results of images change when the image plane is far from the observer, AMFF-Net scales the AGI and feeds the scaled image and the original image into the image encoder of the pre-trained Contrastive Language-Image Pretraining (CLIP) model. Subsequently, the extracted multi-scale features are fused by the Adaptive Feature Fusion (AFF) block. The fused features contain information from different scales of the image, so they can more effectively characterize the distortion and authenticity of the AGI. For the prediction of content consistency, AMFF-Net encodes the textual cues using the text encoder in the pre-trained CLIP and calculates the similarity between the obtained textual features and the fused multi-scale features.

[0045] The model framework constructed in this embodiment is as follows Figure 1As shown in the figure, firstly, the images and text prompts of the AGI dataset are obtained, and then the images are scaled to obtain images with 1.5 times and 0.5 times the resolution. The text prompts and three images of different scales are input into the network. First, the text features are obtained through the pre-trained text encoder. At the same time, the features of the three images are obtained through the image encoder of the CLIP model, and then the three features are fused through the adaptive feature fusion module to obtain multi-scale fusion features. On the one hand, the multi-scale fusion features are passed into two multi-layer perceptrons to obtain the visual quality score and authenticity score respectively, and on the other hand, the cosine similarity with the text features is calculated to obtain the content consistency score.

[0046] (1) Selection of test platform: In order to verify the effectiveness of the method, the AMFF-Net model of this embodiment is implemented under the open source PyTorch deep learning framework, and trained and tested on a server equipped with NVIDIA GeForce RTX 3090 GPU and Intel Xeon Silver 4210R CPU.

[0047] (2) Dataset acquisition: The AI-generated images used in this embodiment are derived from three publicly available AI-generated image quality assessment datasets, including AGIQA-3K, AIGCIQA2023, and P KU-I2IQA. AGIQA-3K contains 2,982 AI-generated images from six text-based image models, namely GLIDE, Stable DiffusionV1.5, Stable Diffusion XL2.2, At tnGAN, Midjourney, and DALLE2. AGIQA-3K also provides 300 text prompts of different scenes and styles, and each image contains visual quality and consistency scores. AIGCIQA2023 consists of 2,400 AI-generated images, which are generated by 6 text-based image models using 100 text prompts, including GLIDE, Lafite, Stable Diffusion, DALLE, Unidiffusion, and Controlnet. For each image, the database provides visual quality, authenticity, and consistency scores. PKU-I2IQA contains 1,600 generated images generated by the Midjourney and Stable Diffusion V1.5 image-generated models. In addition, the database also provides 200 text prompts of different scenes and styles, and the labels of each image include visual quality, authenticity, and consistency scores.

[0048] This embodiment proposes a simple and effective multi-task image quality assessment (IQA) network, named AMFF-Net, for AI generated images (AGI). The design motivations of AMFF-Net are as follows: 1) Inspired by the fact that the perceptibility of image details is highly dependent on the distance between the image plane and the observer, AMFF-Net scales up and down AGI and encodes the scaled and original size images to capture image details at different granularity levels, which is closer to the subtle and holistic characteristics of human visual perception. 2) Through multi-scale image representation, an adaptive feature fusion block is used to adaptively combine image details at different resolutions, which helps accurate quality and authenticity prediction. 3) AMFF-Net evaluates AGI from multiple dimensions, namely visual quality, authenticity, and content consistency, aiming to help users understand the quality of AGIs more comprehensively than current methods that only provide one-dimensional predictions. In addition, calculating the similarity between text features and image features, without the need for designer experience, is a good choice for evaluating the consistency between prompts and generated images.

[0049] Figure 1 Describes the overall architecture of AMFF-Net. AMFF-Net takes textual cues and multi-scale AGI as input and outputs multi-dimensional quality scores, including a content consistency score S C , visual quality score S V and authenticity score S A .

[0050] Overall architecture: Specifically, considering that global and local details affect the subjective evaluation scores of visual quality and authenticity, AMFF-Net inputs multi-scale AGI images, namely I 1.5× ,I 1.0× and I 0.5× , representing 1.5 times, 1.0 times, and 0.5 times the resolution of the original image, respectively. These images are input into the image encoder in the pre-trained CLIP model to obtain multi-scale semantic representations, denoted as and In this embodiment, ResNet50 is selected as the image encoder due to its widely verified effectiveness in image quality assessment tasks. The semantic features of images of different scales are fused by an AFF module to form fused features. Next, F I It is fed into two parallel multi-layer perceptrons (MLPs), which consist of two fully connected layers to predict visual quality and authenticity respectively. The number of neural nodes in MLP is {1024, 256, 1}. For content consistency prediction, the Transformer-based text encoder of CLIP model is first used, and then the extracted text features F are calculated. T And fusion image features FI The cosine similarity of is used to predict the consistency score.

[0051] Adaptive feature fusion module: In previous work, a widely used method is to directly aggregate or add features of different scales. However, such a method can easily cause information of different scales to be masked, making it unsuitable for image quality assessment tasks. In this study, this embodiment proposes a new feature fusion module, called AFF, to adaptively fuse these features.

[0052] Figure 2 The structure of the AFF module is shown. The image encoder obtains semantic features at three scales, namely and Passed to the AFF module. First, these multi-scale features are stacked together to form Next, V I It is processed by two linear layers and a Softmax function. Between the two linear layers, a ReLU activation function is embedded. Next, the processing result is processed by a split operation to obtain three independent features, namely and Represents the weights associated with the three image scales. Using multi-scale semantic features and and their scale-specific weights and The fusion feature F is obtained by element-wise multiplication and element-wise addition operations. I :

[0053]

[0054] This process ensures the adaptive integration of multi-scale semantic features.

[0055] (4) Loss function: AMFF-Net evaluates the quality of AGI in three dimensions. The total loss function L for training AMFF-Net is the linear addition of the three parts:

[0056] L=L C +L V +L A

[0057] L C , L V and L A They are the loss functions used for the three tasks, namely content consistency score prediction, visual quality score prediction, and authenticity score prediction. For the content consistency score prediction task, this embodiment uses the fidelity loss function:

[0058]

[0059] Where N is the number of images in a mini-batch, P i,j is a binary function comparing the i-th image and the j-th image. If P i,j =1; otherwise, P i,j = 0. Here, Q V is the true score of content consistency. is the probability that the i-th image is better than the j-th image calculated by the Thurstone model. For the visual quality and authenticity score prediction tasks, the mean square error loss is used as the loss function to train the network:

[0060]

[0061]

[0062] Where N is the number of images in a mini-batch, and Represent the true quality score and authenticity score of the nth image respectively, and are the quality score and authenticity score of the network prediction, respectively.

[0063] (5) Network training: The AMFF-Net proposed in this embodiment is implemented through the open source Pytorch library. The data set is divided into a data set and a test set in a ratio of 8:2. First, the text prompts and images of the training set are obtained, and then processed to obtain images with 1.5 times the resolution and 0.5 times the resolution. The scaled image, the original image and the text prompt are input into the network to obtain the prediction score, calculate the loss, perform back propagation training, and then adjust the parameters by the gradient descent method to reduce the loss value. Repeat the above process for multiple trainings.

[0064] Specifically, during the network training process, the AdamW optimizer is used to adjust the parameters, and the batch size is set to 32. The model is trained in an end-to-end manner, with 120 training cycles set, and these 120 cycles are divided into two stages. In the first stage, the parameters of the image encoder and text encoder modules of the CLIP model are frozen to train the rest of the AMFF-Net. For the AIGCI QA2023 dataset, the learning rate is set to 1×10 -3 , the learning rate of the other two datasets is set to 5×10 -4 This phase is limited to the first 20 cycles. In the second phase, the text and image encoders of the CLIP model are unfrozen and 5×10 -6 The learning rate is adjusted to 5×10 -7In addition, this embodiment uses an early stopping strategy. When the model performance does not improve for 20 consecutive cycles, the training stops and the weight parameters when the network performance is optimal are saved.

[0065] This embodiment uses Pearson linear correlation coefficient (PLCC), Spearman rank correlation coefficient (SRCC) and Kendall rank correlation coefficient (KRCC) as performance indicators. Among them, PLCC measures the linear correlation between the predicted scores and the actual values ​​collected from the subjective ratings. SRCC and KRCC calculate the ranking correlation between the predicted scores and the actual values. The higher the PLCC, SRCC and KRCC, the better the performance of the network. According to the recommendations of VQEG, before calculating PLCC, the predicted scores are mapped using a four-parameter logistic function:

[0066]

[0067] Among them is The mapping values ​​of the predicted scores s, κ1~κ4 can be obtained by comparing the predicted values ​​with their corresponding true values ​​using the least squares method.

[0068] (6) Network testing: The parameters that obtained the best indicators during training are used as the weight parameters of the MAFF-Net model, and the images of the test set are tested in this network. First, the images are also scaled to obtain images with 1.5 times and 0.5 times resolution. These two images, as well as the original image and text prompts, are input into the network to generate prediction scores for image consistency, visual quality, and authenticity. PLCC, SRCC, and KRCC of the entire test set are calculated as performance indicators of the model.

[0069] (7) Method performance: The PLCC, SRCC, and KRCC results of the test set demonstrate the superiority of the network model designed in this embodiment. Compared with other image quality evaluation algorithms, this algorithm has achieved the most optimal PLCC, SRCC, and KRCC indicators in terms of content consistency, authenticity, and visual quality in many data sets.

[0070] An image quality evaluation system for AI images, comprising:

[0071] A data acquisition module is used to obtain an initial AI image and corresponding text prompt data; pre-process the initial AI image to obtain a number of AI images of different scales;

[0072] An image quality assessment module is used to input AI images of different scales and corresponding text prompt data into a multi-scale image quality assessment model for quality assessment to obtain an image quality assessment score; wherein the image quality assessment score includes a content consistency score, a visual quality score and an authenticity score of the AI ​​image; the multi-scale image quality assessment model includes a feature extraction module, an AFF module and an evaluation module connected in sequence; the feature extraction module is constructed based on an image encoder, and the evaluation module includes a number of fully connected layers arranged in parallel.

[0073] An electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program so that the electronic device executes the image quality assessment method of an AI image.

[0074] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the image quality evaluation method of an AI image is implemented.

[0075] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An image quality assessment method for AI images, It is characterized in that include: Get the initial AI image and corresponding text prompt data; Preprocessing the initial AI image to obtain a plurality of AI images of different scales; Inputting AI images of different scales and corresponding text prompt data into a multi-scale image quality assessment model for quality assessment to obtain an image quality assessment score; wherein, the image quality assessment score includes the content consistency score, visual quality score and authenticity score of the AI ​​image; the multi-scale image quality assessment model includes a feature extraction module, an AFF module and an evaluation module connected in sequence; the feature extraction module is constructed based on an image encoder, and the AFF module includes a feature stacking layer, a first linear layer, a second linear layer and a feature fusion layer connected in sequence, a ReLU activation function is inserted between the first linear layer and the second linear layer, and a Softmax function is set after the second linear layer; the evaluation module includes a number of fully connected layers set in parallel; The training process of the multi-scale image quality assessment model includes: Acquire training data; the training data includes AI training images and corresponding image quality evaluation scores; construct an initial multi-scale image quality evaluation model, input the training data into the initial multi-scale image quality evaluation model for quality evaluation, and train with the goal of minimizing the loss between the initial training result after quality evaluation and the AI ​​training image and the corresponding image quality evaluation score, so as to obtain the multi-scale image quality evaluation model; The processing process of the multi-scale image quality assessment model includes: Inputting AI images of different scales and corresponding text prompt data into the feature extraction module to perform multi-scale feature extraction, thereby obtaining multi-scale semantic feature data and text feature data; Inputting the multi-scale semantic feature data into the AFF module for adaptive feature fusion to obtain fused picture feature data; Inputting the fused image feature data into the evaluation module for score prediction to obtain a visual quality score and authenticity score corresponding to the initial AI image; The content consistency score corresponding to the initial AI image is calculated based on the fused image feature data and the text feature data.

2. The method for evaluating the image quality of an AI image according to claim 1, It is characterized in that The labels of the initial AI images all include a visual quality score, a authenticity score, and a consistency score.

3. The method for evaluating the image quality of an AI image according to claim 1, It is characterized in that Preprocessing the initial AI image to obtain a preprocessed initial AI image specifically includes: The initial AI image is scaled to obtain a plurality of AI images of different scales corresponding to the initial AI image.

4. An image quality assessment system for AI images, It is characterized in that include: A data acquisition module is used to obtain an initial AI image and corresponding text prompt data; pre-process the initial AI image to obtain a number of AI images of different scales; An image quality assessment module is used to input AI images of different scales and corresponding text prompt data into a multi-scale image quality assessment model for quality assessment to obtain an image quality assessment score; wherein the image quality assessment score includes a content consistency score, a visual quality score and a authenticity score of the AI ​​image; the multi-scale image quality assessment model includes a feature extraction module, an AFF module and an assessment module connected in sequence; the feature extraction module is constructed based on an image encoder, and the AFF module includes a feature stacking layer, a first linear layer, a second linear layer and a feature fusion layer connected in sequence, a ReLU activation function is inserted between the first linear layer and the second linear layer, and a Softmax function is set after the second linear layer; the evaluation module includes a plurality of fully connected layers arranged in parallel; The training process of the multi-scale image quality assessment model includes: Acquire training data; the training data includes AI training images and corresponding image quality evaluation scores; construct an initial multi-scale image quality evaluation model, input the training data into the initial multi-scale image quality evaluation model for quality evaluation, and train with the goal of minimizing the loss between the initial training result after quality evaluation and the AI ​​training image and the corresponding image quality evaluation score, so as to obtain the multi-scale image quality evaluation model; The processing process of the multi-scale image quality assessment model includes: Inputting AI images of different scales and corresponding text prompt data into the feature extraction module to perform multi-scale feature extraction, thereby obtaining multi-scale semantic feature data and text feature data; Inputting the multi-scale semantic feature data into the AFF module for adaptive feature fusion to obtain fused picture feature data; Inputting the fused image feature data into the evaluation module for score prediction to obtain a visual quality score and authenticity score corresponding to the initial AI image; The content consistency score corresponding to the initial AI image is calculated based on the fused image feature data and the text feature data.

5. An electronic device, It is characterized in that It includes a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform an image quality assessment method for an AI image according to any one of claims 1-3.

6. A computer-readable storage medium, It is characterized in that It stores a computer program, which, when executed by a processor, implements an image quality assessment method for an AI image as described in any one of claims 1 to 3.