Image quality evaluation method and device based on multi-submodel collaboration, equipment and medium

Abstract

The embodiment of the invention discloses an image quality evaluation method and device based on multi-submodel collaboration, equipment and a medium, and the method comprises the steps: carrying out the feature extraction of to-be-evaluated image data, and obtaining an image feature vector and a context feature vector of the image data; splicing the image feature vector and the context feature vector by a dynamic gating network to form a fusion feature vector, and calculating model weights of the plurality of sub-models according to the fusion feature vector; selecting at least one target sub-model from the plurality of sub-models according to the model weight, and activating the at least one target sub-model to perform image quality evaluation from the corresponding image quality dimension; and integrating output results of the at least one target sub-model to obtain a target image quality evaluation result. According to the image quality evaluation scheme, low-order image quality degradation and high-order perception attributes can be covered at the same time, dynamic distribution of the model weights of the sub-models is achieved, the sub-models can be activated as needed, and the image quality evaluation efficiency is improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of Internet technology, and specifically to a method, apparatus, device, and medium for image quality assessment based on multi-sub-model collaboration. Background Art

[0002] In the existing technology, image quality assessment technology mainly relies on traditional mathematical indicators and end-to-end models based on deep learning. Among them, traditional mathematical indicators include peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Traditional mathematical indicators generate objective scores by calculating pixel-level differences or structural similarities. Although they have clear physical meanings, their limitations are that they can only reflect low-level image quality degradation such as noise and blur, and cannot model the nonlinear and complex perception of the human visual system for high-level image quality attributes such as dynamic range and color naturalness, resulting in a deviation between the score and the subjective experience. In addition, mathematical indicators such as peak signal-to-noise ratio and structural similarity rely on manually designed mathematical formulas, and their calculation rules are rigid, making it difficult to adapt to the image quality requirements of new imaging technologies (such as computational photography, neural rendering, etc.).

[0003] In recent years, single-model methods based on deep learning have improved evaluation accuracy to a certain extent through data-driven methods. However, their single network structure is limited by the solidification of feature representation capabilities and is difficult to cover multi-dimensional image quality attributes at the same time. When a single network structure is used to learn multiple image quality dimensions at the same time, the gradient directions of different tasks interfere with each other, which may lead to unstable model convergence. Summary of the Invention

[0004] In view of the above problems, this application proposes a method, device, equipment and medium for image quality assessment based on multi-submodel collaboration to solve the following problems: existing image quality assessment methods are difficult to accurately assess from the perspective of human visual system perception.

[0005] According to one aspect of an embodiment of the present application, a method for image quality assessment based on multi-sub-model collaboration is provided, comprising:

[0006] Perform feature extraction on the image data to be evaluated to obtain an image feature vector and a context feature vector of the image data;

[0007] The dynamic gating network concatenates the image feature vector and the context feature vector to form a fused feature vector. The model weights of multiple sub-models are calculated based on the fused feature vector. Different sub-models are used to evaluate image quality from different image quality dimensions.

[0008] Selecting at least one target sub-model from the multiple sub-models according to the model weight, and activating the at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension;

[0009] The output results of at least one target sub-model are integrated to obtain a target image quality evaluation result.

[0010] Furthermore, performing feature extraction on the image data to be evaluated to obtain an image feature vector and a context feature vector of the image data further includes:

[0011] Calculate based on the image data to be evaluated to obtain basic statistical features of the image data;

[0012] Extract semantic features of image data through neural networks;

[0013] Using the image feature processing network branch, the basic statistical features and semantic features are processed to obtain the image feature vector of the image data;

[0014] The context feature processing network branch is used to process the basic statistical features and semantic features to obtain the context feature vector of the image data.

[0015] Furthermore, before the dynamic gating network concatenates the image feature vector and the context feature vector to form a fused feature vector, the method further includes:

[0016] Obtaining object interaction data associated with the image data, performing image quality assessment preference analysis on the object interaction data, and obtaining an object preference vector;

[0017] The dynamic gating network concatenates the image feature vector and the context feature vector to form a fused feature vector:

[0018] The dynamic gating network concatenates the image feature vector, context feature vector, and object preference vector to form a fused feature vector.

[0019] Furthermore, calculating the model weights of the multiple sub-models according to the fused feature vector further includes:

[0020] The fused feature vector is processed using a fully connected neural network and a multi-classification activation function to obtain the model weights of multiple sub-models.

[0021] Furthermore, selecting at least one target sub-model from the plurality of sub-models according to the model weights further includes:

[0022] The model weight of each sub-model is compared with the activation weight threshold, and the sub-model whose model weight is greater than the activation weight threshold is determined as the target sub-model.

[0023] Furthermore, activating at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension further includes:

[0024] Activate at least one target sub-model, input the image feature vector and the context feature vector into the at least one target sub-model, and have the at least one target sub-model perform image quality evaluation from the corresponding image quality dimension based on the image feature vector and the context feature vector to obtain the output result of the at least one target sub-model.

[0025] Furthermore, performing image quality evaluation from corresponding image quality dimensions according to the image feature vector and the context feature vector by at least one target sub-model further includes:

[0026] During the image quality assessment process, different target sub-models exchange intermediate feature information through the attention module.

[0027] Furthermore, activating at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension further includes:

[0028] When at least one target sub-model includes a target sub-model at the first level and a target sub-model at the second level, the target sub-model at the first level is activated first to perform image quality evaluation from the corresponding image quality dimension. If the output result of the target sub-model at the first level meets the basic evaluation requirements, the target sub-model at the second level is activated to perform image quality evaluation from the corresponding image quality dimension.

[0029] Furthermore, integrating the output results of at least one target sub-model to obtain a target image quality assessment result further includes:

[0030] According to the model weight of at least one target sub-model, the image quality dimension score in the output result of at least one target sub-model is weightedly calculated to obtain a weighted total score, and the weighted total score is normalized to form the target image quality score in the target image quality assessment result.

[0031] According to another aspect of an embodiment of the present application, a device for image quality assessment based on multi-sub-model collaboration is provided, comprising:

[0032] A feature extraction module, adapted to extract features from the image data to be evaluated, and obtain an image feature vector and a context feature vector of the image data;

[0033] A gating module, adapted to concatenate the image feature vector and the context feature vector using a dynamic gating network to form a fused feature vector, and to calculate the model weights of multiple sub-models based on the fused feature vector; wherein different sub-models are used to evaluate image quality from different image quality dimensions;

[0034] a processing module, adapted to select at least one target sub-model from the plurality of sub-models according to a model weight, and activate the at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension;

[0035] The result integration module is adapted to integrate the output results of at least one target sub-model to obtain a target image quality assessment result.

[0036] Furthermore, the feature extraction module is further adapted to:

[0037] Calculate based on the image data to be evaluated to obtain basic statistical features of the image data;

[0038] Extract semantic features of image data through neural networks;

[0039] Using the image feature processing network branch, the basic statistical features and semantic features are processed to obtain the image feature vector of the image data;

[0040] The context feature processing network branch is used to process the basic statistical features and semantic features to obtain the context feature vector of the image data.

[0041] Furthermore, the feature extraction module is further adapted to:

[0042] Obtaining object interaction data associated with the image data, performing image quality assessment preference analysis on the object interaction data, and obtaining an object preference vector;

[0043] The gate control module is further adapted to:

[0044] The dynamic gating network concatenates the image feature vector, context feature vector, and object preference vector to form a fused feature vector.

[0045] Furthermore, the gate control module is further adapted to:

[0046] The fused feature vector is processed using a fully connected neural network and a multi-classification activation function to obtain the model weights of multiple sub-models.

[0047] Furthermore, the processing module is further adapted to:

[0048] The model weight of each sub-model is compared with the activation weight threshold, and the sub-model whose model weight is greater than the activation weight threshold is determined as the target sub-model.

[0049] Furthermore, the processing module is further adapted to:

[0050] Activate at least one target sub-model, input the image feature vector and the context feature vector into the at least one target sub-model, and have the at least one target sub-model perform image quality evaluation from the corresponding image quality dimension based on the image feature vector and the context feature vector to obtain the output result of the at least one target sub-model.

[0051] Furthermore, the processing module is further adapted to:

[0052] During the image quality assessment process, different target sub-models exchange intermediate feature information through the attention module.

[0053] Furthermore, the processing module is further adapted to:

[0054] When at least one target sub-model includes a target sub-model at the first level and a target sub-model at the second level, the target sub-model at the first level is activated first to perform image quality evaluation from the corresponding image quality dimension. If the output result of the target sub-model at the first level meets the basic evaluation requirements, the target sub-model at the second level is activated to perform image quality evaluation from the corresponding image quality dimension.

[0055] Furthermore, the result integration module is further adapted to:

[0056] According to the model weight of at least one target sub-model, the image quality dimension score in the output result of at least one target sub-model is weightedly calculated to obtain a weighted total score, and the weighted total score is normalized to form the target image quality score in the target image quality assessment result.

[0057] According to another aspect of an embodiment of the present application, a computing device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus;

[0058] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the above-mentioned image quality assessment method based on multi-sub-model collaboration.

[0059] According to another aspect of the embodiments of the present application, a computer storage medium is provided, in which at least one executable instruction is stored. The executable instruction enables a processor to perform operations corresponding to the above-mentioned image quality assessment method based on multi-sub-model collaboration.

[0060] According to another aspect of an embodiment of the present application, a computer program product is provided, comprising at least one executable instruction, wherein the executable instruction enables a processor to perform operations corresponding to the above-mentioned image quality assessment method based on multi-sub-model collaboration.

[0061] According to the technical solution provided in the embodiment of the present application, through the sub-model division of labor design, image quality evaluation is performed separately for each image quality dimension, which effectively avoids multi-task conflicts and effectively improves the evaluation accuracy of a single image quality dimension, so that the image quality evaluation can simultaneously cover low-level image quality degradation such as noise and blur, as well as high-level perceptual attributes such as dynamic range and color naturalness, and can more accurately and comprehensively reflect the perceptual ability of the human visual system, solving the shortcomings of existing image quality evaluation methods in the perception of the human visual system; the dynamic gating network can dynamically allocate model weights of different sub-models according to the image feature vector and context feature vector of the image data, ensuring that the image quality evaluation process can focus well on the key image quality dimensions of the current image data to be evaluated, activate sub-models on demand, effectively reduce redundant calculations, improve image quality evaluation efficiency, and realize fast and automated image quality evaluation, which can well adapt to the image quality requirements of new imaging technologies such as computational photography and neural rendering.

[0062] The above description is only an overview of the technical solution of the embodiment of the present application. In order to more clearly understand the technical means of the embodiment of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the embodiment of the present application more obvious and easy to understand, the specific implementation method of the embodiment of the present application is specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the embodiments of the present application. The same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0064] Figure 1 A schematic diagram of a process for image quality assessment based on multi-sub-model collaboration according to an embodiment of the present application is shown;

[0065] Figure 2 A schematic diagram of a process for image quality assessment based on multi-sub-model collaboration according to another embodiment of the present application is shown;

[0066] Figure 3 The following is a structural block diagram of an image quality assessment device based on multi-sub-model collaboration according to an embodiment of the present application;

[0067] Figure 4 A schematic structural diagram of a computing device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0068] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0069] First, the terms involved in one or more embodiments of the present application are explained.

[0070] Mixture of Experts (MoE): A machine learning architecture composed of multiple sub-models (i.e., "experts") that selects and weights the outputs of different sub-models through a gating network. The final result is determined by the weighted combination of the outputs of each sub-model.

[0071] Dynamic gating networks (DGNs) are neural network modules that dynamically adjust the model weights of each sub-model in real time based on input features (e.g., image content, device parameters, etc.) in this application. Compared to existing static gating networks, DGNs offer greater adaptability and interpretability.

[0072] Sub-model: refers to the expert network within the hybrid expert model that focuses on a specific task. In the field of image quality assessment in the embodiments of this application, each sub-model can be designed to perform image quality assessment on a single image quality dimension (e.g., noise, color naturalness, contrast, dynamic range, etc.).

[0073] Multilayer Perceptron (MLP): A feedforward artificial neural network consisting of multiple fully connected layers. It can learn complex nonlinear feature mappings and is widely used in tasks such as classification, regression, and feature extraction.

[0074] Cross-attention mechanism: refers to a module that allows information to be exchanged between different sub-models, and calculates feature correlation through attention weights (for example, the feature interaction between the sub-model corresponding to the color naturalness dimension and the sub-model corresponding to the contrast dimension).

[0075] Hierarchical gating: This refers to a mechanism for selecting sub-models in stages, such as global evaluation followed by local refinement, or triggering different sub-models based on the severity of image quality issues.

[0076] Figure 1 FIG. 1 shows a flow chart of a method for image quality assessment based on multi-sub-model collaboration according to an embodiment of the present application. Figure 1 As shown, the method includes the following steps:

[0077] Step S101 : performing feature extraction on the image data to be evaluated to obtain an image feature vector and a context feature vector of the image data.

[0078] The mixture of experts (MoE) model has demonstrated the potential of multiple sub-models to collaborate in natural language processing and image classification tasks, which can improve overall model performance. However, its general framework has significant flaws in direct application in the field of image quality assessment. Among them, the gating network in the existing mixture of experts model only relies on low-level, shallow features of the input image and does not incorporate imaging context information, resulting in the sub-model selection being disconnected from the physical causes of image quality problems. In addition, the existing mixture of experts model lacks a sub-model division of labor design for image quality assessment. The general sub-model usually focuses on semantic feature extraction rather than fine-grained image quality dimensions.

[0079] In an embodiment of the present application, an improved architecture based on a hybrid expert model is introduced into the field of image quality assessment. Specifically, a dynamic gating network and a sub-model group are set up, and the input features of the dynamic gating network are enriched. The sub-model group contains multiple independent sub-models, and different sub-models are used to perform image quality assessment from different image quality dimensions. Through the sub-model division of labor design, image quality assessment is performed separately for image quality dimensions such as noise, blur, color naturalness, contrast, and dynamic range, avoiding multi-task conflicts and effectively improving the assessment accuracy of a single image quality dimension, so that the image quality assessment can simultaneously cover low-level image quality degradation such as noise and blur, as well as high-order perceptual attributes such as dynamic range and color naturalness.

[0080] In step S101, image data to be evaluated is obtained and feature extraction is performed on the image data to obtain an image feature vector and a context feature vector of the image data. Specifically, basic statistical features and semantic features of the image data can be first extracted from the image data. Basic statistical features are low-level, shallow features, while semantic features are high-level, deep features. Then, the basic statistical features and semantic features are fused using network branches to obtain the image feature vector and context feature vector of the image data.

[0081] Among them, both the image feature vector and the context feature vector are integrated with low-level features and high-level features. The image feature vector is a feature vector used to reflect the global semantic information and overall image quality characteristics of the image data, while the context feature vector is a feature vector used to reflect the image content and its context-related information (such as spatial context, semantic context, temporal context, scene context, etc.).

[0082] In step S102 , the dynamic gating network concatenates the image feature vector and the context feature vector to form a fused feature vector, and calculates the model weights of multiple sub-models based on the fused feature vector.

[0083] The dynamic gating network dynamically assigns model weights to different sub-models based on the image feature vector and context feature vector of the image data, ensuring that the image quality assessment process is well focused on the key image quality dimensions of the image data being evaluated. To facilitate the calculation of the model weights of each sub-model, the dynamic gating network concatenates the image feature vector and the context feature vector, thereby concatenating the multiple input feature vectors into a unified vector, called a fused feature vector. This unified vector is then processed using a fully connected neural network and a multi-class activation function to calculate the model weights of the multiple sub-models.

[0084] Among them, the model weight of the sub-model refers to the weight coefficient generated by the dynamic gating network for dynamically allocating input data to different sub-models, which reflects the contribution and matching degree of each sub-model to the current input data. The model weight of a single sub-model can be between 0 and 1. The higher the model weight of a sub-model, the greater the influence of the sub-model on the output. Different sub-models are used to evaluate image quality from different image quality dimensions. There is a one-to-one correspondence between the sub-model and the image quality dimension, and each sub-model focuses on the evaluation of a single image quality dimension. Image quality dimensions may include: noise dimension, blur detection dimension, color naturalness dimension, contrast dimension, dynamic range dimension, artifact dimension, etc. Image quality dimensions may also include other dimensions, which are not specifically limited here.

[0085] Step S103 : selecting at least one target sub-model from the plurality of sub-models according to the model weights, and activating the at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension.

[0086] After calculating the model weights of multiple sub-models, the model weight of each sub-model can be compared with the activation weight threshold to select which sub-model or sub-models need to be activated for the current image data to be evaluated. In this embodiment of the present application, the sub-model to be activated is called the target sub-model. The number of target sub-models can be one or more. After determining at least one target sub-model, at least one target sub-model can be activated to perform image quality evaluation from its corresponding image quality dimension. The embodiment of the present application realizes dynamic decoupling of image quality dimensions based on the hybrid expert model architecture, activates sub-models on demand based on image content perception, avoids interference from irrelevant tasks, and helps reduce redundant calculations through sparse activation, which can meet real-time requirements.

[0087] For example, the sub-model group includes a sub-model corresponding to the noise dimension, a sub-model corresponding to the color naturalness dimension, a sub-model corresponding to the contrast dimension, a sub-model corresponding to the dynamic range dimension, and a sub-model corresponding to the artifact dimension. If the determined target sub-model includes a sub-model corresponding to the dynamic range dimension, a sub-model corresponding to the noise dimension, and a sub-model corresponding to the color naturalness dimension, then the sub-model corresponding to the dynamic range dimension is activated to evaluate the image quality from the dynamic range dimension, the sub-model corresponding to the noise dimension is activated to evaluate the image quality from the noise dimension, and the sub-model corresponding to the color naturalness dimension is activated to evaluate the image quality from the color naturalness dimension.

[0088] Step S104: integrating the output results of at least one target sub-model to obtain a target image quality assessment result.

[0089] After obtaining the output results of at least one target sub-model, the output results of at least one target sub-model are integrated, such as by weighted score calculation, image quality description information summary, and restoration suggestion summary, to obtain a final target image quality assessment result for the image data. The target image quality assessment result may include image quality dimension scores for each image quality dimension, an overall target image quality score after integrating multiple image quality dimensions, image quality description information for each image quality dimension, and restoration suggestions integrated across multiple image quality dimensions.

[0090] According to the image quality assessment method based on multi-sub-model collaboration provided by the embodiment of the present application, through the sub-model division of labor design, image quality assessment is performed separately for each image quality dimension, effectively avoiding multi-task conflicts and effectively improving the assessment accuracy of a single image quality dimension, so that the image quality assessment can simultaneously cover low-level image quality degradation such as noise and blur, as well as high-level perceptual attributes such as dynamic range and color naturalness, and can more accurately and comprehensively reflect the perceptual ability of the human visual system, solving the shortcomings of existing image quality assessment methods in the perception of the human visual system; the dynamic gating network can dynamically allocate model weights of different sub-models according to the image feature vector and context feature vector of the image data, ensuring that the image quality assessment process can well focus on the key image quality dimensions of the current image data to be evaluated, activating sub-models on demand, effectively reducing redundant calculations, improving image quality assessment efficiency, and realizing fast and automated image quality assessment, which can well adapt to the image quality requirements of new imaging technologies such as computational photography and neural rendering.

[0091] Figure 2 FIG. 4 shows a flow chart of a method for image quality assessment based on multi-sub-model collaboration according to another embodiment of the present application. Figure 2 As shown, the method includes the following steps:

[0092] Step S201 : performing feature extraction on the image data to be evaluated to obtain an image feature vector and a context feature vector of the image data.

[0093] Acquire the image data to be evaluated. The embodiments of the present application support image quality evaluation of image data of various file formats, data sources, and pixel sizes. For example, the image data can be a frame image in a video, or a photo uploaded by a user using various devices, or an image obtained from the Internet. Parse the image data and perform feature extraction on the image data. The image data may include image metadata, image content, and other data. For example, the metadata may include file format, file size, creation time, shooting parameters (such as aperture, shutter speed, camera sensor sensitivity, camera model, etc.), color space, bit depth, pixel size, etc.

[0094] Specifically, calculations are performed based on the image data to be evaluated to obtain basic statistical features of the image data. Basic statistical features refer to features that can be obtained by directly performing mathematical calculations on the image data, such as brightness histogram features, noise level features, edge sharpness features, etc. Basic statistical features are low-level features; semantic features of the image data are extracted through a neural network. Semantic features are high-level features. Semantic features can specifically be feature maps obtained after processing by a neural network, etc.

[0095] In the embodiment of the present application, both the image feature vector and the context feature vector of the image data are obtained by fusing basic statistical features and semantic features. In order to facilitate the acquisition of the image feature vector and the context feature vector, an image feature processing network branch and a context feature processing network branch are constructed respectively.

[0096] The basic statistical features and semantic features are input into the image feature processing network branch, which processes the basic statistical features and semantic features to obtain an image feature vector for the image data. The image feature processing network branch can be constructed using a lightweight convolutional neural network to process the basic statistical features and semantic features, extracting global semantic information from the image data, including texture, contrast, color distribution, and overall image quality characteristics. For example, this identifies the scene classification to which the image data belongs, which may include night scenes, portrait scenes, and so on. The image feature vector is obtained through the above processing.

[0097] Basic statistical and semantic features are input into the contextual feature processing network branch, which processes the basic statistical and semantic features to obtain a contextual feature vector for the image data. The contextual feature processing network branch may include a multi-layer perceptron (MLP). The MLP normalizes the basic statistical and semantic features and then feeds them into the MLP for encoding. The MLP analyzes the image content and its contextual associations, quantifies physical indicators of specific image quality issues (e.g., highlight overexposure ratio, noise intensity, etc.), and obtains a contextual feature vector for the image data. This contextual feature vector serves as an objective basis for guiding sub-model selection.

[0098] The dimensions of the image feature vector and the context feature vector may be the same or different, and are not specifically limited here. For example, the image feature vector may be 512-dimensional, and the context feature vector may be 128-dimensional.

[0099] Step S202 : acquiring object interaction data associated with the image data, performing image quality assessment preference analysis on the object interaction data, and obtaining an object preference vector.

[0100] Different subjects have different preferences for image quality assessment. For example, photography enthusiasts are more interested in dynamic range, while ordinary subjects are more concerned with color naturalness. Object interaction data associated with the image data is obtained. For example, the ID of the subject providing the image data is obtained. Then, based on the object ID, the platform or database is used to query the object interaction data associated with that object ID. This object interaction data may include adjustments to the subject's image quality score and settings for the subject's image quality dimension preferences. Image quality assessment preferences can be analyzed on this object interaction data using methods such as embedded table mapping and neural networks to generate an object preference vector.

[0101] In addition to using the image feature vector and the context feature vector as inputs to the dynamic gating network, the embodiment of the present application also uses the object preference vector as input to the dynamic gating network, thereby achieving personalized adjustment of the model weights of each sub-model and effectively ensuring the comprehensiveness of the sub-model selection.

[0102] In step S203 , the dynamic gating network concatenates the image feature vector, the context feature vector, and the object preference vector to form a fused feature vector, and calculates the model weights of the multiple sub-models based on the fused feature vector.

[0103] To facilitate the calculation of the model weights of each sub-model, the dynamic gating network concatenates the image feature vector, the context feature vector, and the object preference vector, thereby concatenating the multiple input feature vectors into a unified fused feature vector. For example, when the image feature vector is 512-dimensional, the context feature vector is 128-dimensional, and the object preference vector is 64-dimensional, these three vectors can be concatenated into a 704-dimensional fused feature vector. The fused feature vector can be expressed as F_gate=Concat(F_img, F_meta, F_user)∈704, where F_gate represents the fused feature vector, F_img represents the image feature vector, F_meta represents the context feature vector, and F_user represents the object preference vector.

[0104] After forming the fused feature vector, the fused feature vector is processed using a fully connected neural network and a multi-classification activation function to obtain the model weights of multiple sub-models. In specific applications, the fused feature vector can be processed in sequence through two layers of fully connected neural networks and multi-classification activation functions to generate a model selection weight vector, which contains the model weights of multiple sub-models. The model selection weight vector can be expressed as g = [g1, g2, ..., gN], where g represents the model selection weight vector, and g1 to gN represent the model weights of the first to Nth sub-models. Among them, the multi-classification activation function refers to an activation function suitable for multi-classification problems, such as the Softmax function, the Sigmoid function, etc. Those skilled in the art can select a specific multi-classification activation function according to actual needs, which is not limited here.

[0105] For example, when it is known from the fused feature vector that the image data is a high-sensitivity night scene, the model weight of the sub-model corresponding to the calculated noise dimension will be significantly higher than that of other sub-models.

[0106] In the embodiments of this application, the dynamic gating network calculates model weights for multiple sub-models, directly linking them to image feature vectors, context feature vectors, and object preference vectors. This allows the network to automatically adjust the focus of image quality assessment based on different image content perceptions and object preferences, avoiding static assessment bias and supporting visual interpretation. By activating sub-models on demand through the dynamic gating network, interference from irrelevant tasks is effectively avoided. Experimental verification shows that sparse activation can reduce redundant computation by over 70%, effectively meeting real-time requirements.

[0107] Step S204 : selecting at least one target sub-model from the plurality of sub-models according to the model weights, and activating the at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension.

[0108] The embodiment of the present application constructs a sub-model group, which contains multiple independent sub-models. Different sub-models are used to evaluate image quality from different image quality dimensions, and each sub-model processes input data in parallel. The sub-model group realizes accurate evaluation of multiple image quality dimensions through specialized division of labor and coordination mechanism. Among them, there is a one-to-one correspondence between the sub-model and the image quality dimension, and the image quality dimension may include: noise dimension, blur detection dimension, color naturalness dimension, contrast dimension, dynamic range dimension, artifact dimension, etc. Through the sub-model division of labor design, each image quality dimension is evaluated separately to avoid multi-task conflicts. Experimental verification shows that the evaluation accuracy of a single image quality dimension is improved by at least 22%.

[0109] Each sub-model possesses the dual characteristics of task specificity and feature sharing. Specifically, each sub-model focuses on evaluating a single image quality dimension, with its network structure and loss function optimized for that quality dimension. All sub-models share the underlying feature extraction layer—the image feature vector and context feature vector obtained in step S201—effectively reducing duplicate computations and improving resource utilization. Experimental verification shows that sharing the underlying feature extraction layer can reduce duplicate computations by over 60%.

[0110] In the embodiment of the present application, each sub-model is task-adapted. Each sub-model, in addition to sharing the underlying feature extraction layer, is also equipped with a dedicated processing module (i.e., a dedicated evaluation head) to optimize its corresponding image quality dimension. For example, the sub-model corresponding to the noise dimension extracts high-frequency components from the luminance channel (e.g., through wavelet transforms) and analyzes noise distribution patterns, such as Gaussian noise and fixed-pattern noise. The sub-model corresponding to the dynamic range dimension analyzes the distribution characteristics of the luminance histogram, detects overexposed highlights and underexposed shadows, and weights the scores based on the human visual sensitivity curve. The sub-model corresponding to the color naturalness dimension converts the image to the Lab color space and calculates the color gamut coverage and color difference indicators for specific areas. Artifacts refer to abnormal traces or distortions of non-realistic scenes that appear in the image. They are usually introduced by artificial factors such as device hardware, algorithm processing, or transmission compression. Artifacts can reduce image quality and interfere with the accuracy of visual judgment or computer vision tasks. The sub-model corresponding to the artifact dimension detects compression artifacts, such as blocking and ringing, through multi-scale dilated convolution, locates the artifact area, and assesses its severity.

[0111] In an embodiment of the present application, a sub-model is activated on demand through a dynamic gating network. To facilitate the selection of a target sub-model, an activation weight threshold is pre-set, and the corresponding target sub-model is activated according to the activation weight threshold. The model weight of each sub-model is compared with the activation weight threshold to determine whether the model weight is greater than the activation weight threshold. The sub-model with a model weight greater than the activation weight threshold is determined as the target sub-model.

[0112] After determining at least one target sub-model, the at least one target sub-model can be activated to perform image quality assessment based on its corresponding image quality dimension. Specifically, the at least one target sub-model is activated, and the image feature vector and the context feature vector are input to the at least one target sub-model. The at least one target sub-model then performs image quality assessment based on the image feature vector and the context feature vector based on the corresponding image quality dimension, thereby obtaining an output result of the at least one target sub-model.

[0113] In order to further improve the accuracy and efficiency of image quality assessment, a collaborative mechanism between sub-models is also set in the embodiment of the present application. The collaborative mechanism between sub-models includes a cross-attention mechanism and hierarchical gating.

[0114] Regarding the cross-attention mechanism: During the image quality assessment process, different target sub-models exchange intermediate feature information through the attention module, so that the target sub-model can combine the intermediate feature information of other target sub-models to make a comprehensive judgment during the evaluation of the corresponding image quality dimension, effectively improving the assessment accuracy and reducing misjudgments. For example, the sub-model corresponding to the dynamic range dimension passes the brightness distribution characteristics to the sub-model corresponding to the noise dimension to assist it in distinguishing signal-related noise from artifacts caused by uneven lighting; the sub-model corresponding to the color naturalness dimension passes the color difference characteristics to the sub-model corresponding to the contrast dimension to assist it in correcting misjudgments of local contrast (such as contrast perception deviations caused by color casts).

[0115] Specifically, the attention module can be a lightweight attention module. Intermediate feature information refers to feature information generated by a sub-model during the intermediate stages of model processing. Intermediate feature information differs from the sub-model's final output. The cross-attention mechanism can effectively reduce cross-dimensional evaluation errors, for example, by minimizing the impact of color deviation on contrast scores. Experimental verification shows that the cross-attention mechanism reduces cross-dimensional evaluation errors by over 15%.

[0116] For hierarchical gating, sub-models can be divided into different levels, and cascade execution strategies can be set for sub-models belonging to different levels according to the level. For example, the sub-models in the first level may include sub-models corresponding to the noise dimension and the blur detection dimension, and the sub-models in the second level may include sub-models corresponding to the color naturalness dimension, the contrast dimension, the dynamic range dimension, and the artifact dimension.

[0117] When at least one target sub-model includes a target sub-model at the first level and a target sub-model at the second level, the target sub-model at the first level is activated first to perform image quality evaluation from the corresponding image quality dimension. If the output result of the target sub-model at the first level meets the basic evaluation requirements, the target sub-model at the second level is activated to perform image quality evaluation from the corresponding image quality dimension.

[0118] For example, the basic evaluation requirements include a basic scoring threshold, and the target sub-model includes a sub-model corresponding to the noise dimension and the blur detection dimension at the first level, as well as a sub-model corresponding to the dynamic range dimension at the second level. The sub-model corresponding to the noise dimension and the sub-model corresponding to the blur detection dimension are activated first to perform image quality evaluation from the corresponding image quality dimensions; if any of the two sub-models detects serious image quality defects, for example, the image quality dimension score is less than the basic scoring threshold, there is no need to activate the sub-model corresponding to the dynamic range dimension at the second level, and the output results of the sub-model corresponding to the noise dimension and the output results of the sub-model corresponding to the blur detection dimension can be directly integrated to obtain the target image quality evaluation result; if the image quality dimension scores of both sub-models are greater than or equal to the basic scoring threshold, the sub-model corresponding to the dynamic range dimension is activated again. Through the above-mentioned hierarchical gating, resource waste is effectively avoided and the average inference time is reduced. Experimental verification shows that cascade execution can reduce the average inference time by more than 38%.

[0119] Step S205 , integrating the output results of at least one target sub-model to obtain a target image quality assessment result.

[0120] Among them, the target image quality assessment results may include the image quality dimension scores of each image quality dimension, the overall target image quality score after integrating multiple image quality dimensions, the image quality description information of each image quality dimension, and the restoration suggestions after integrating multiple image quality dimensions.

[0121] The target image quality score may be calculated as follows: according to the model weight of at least one target sub-model, the image quality dimension score in the output result of at least one target sub-model is weightedly calculated to obtain a weighted total score, and the weighted total score is normalized to form the target image quality score in the target image quality assessment result.

[0122] Taking a backlit portrait image taken with a mobile phone as an example, the image data to be evaluated is resized to 512×512 during feature extraction. The RAW channel data, image histogram, YUV channel data, and various noise distributions are extracted to obtain the image feature vector and context feature vector. The object preference vector is then analyzed. A dynamic gating network concatenates the image feature vector, context feature vector, and object preference vector to form a fused feature vector. The model weights of multiple sub-models are then calculated based on this fused feature vector.

[0123] Assuming that the calculated model weight of the sub-model corresponding to the dynamic range dimension is a1, the model weight of the sub-model corresponding to the noise dimension is a2, the model weight of the sub-model corresponding to the color naturalness dimension is a3, the model weight of the sub-model corresponding to the artifact dimension is a4, and the preset activation weight threshold is b, then the sub-model whose model weight is greater than the activation weight threshold is determined as the target sub-model. For example, a1, a2 and a3 are all greater than b, then the target sub-model includes the sub-model corresponding to the dynamic range dimension, the model weight of the sub-model corresponding to the noise dimension and the sub-model corresponding to the color naturalness dimension.

[0124] If the output result of the sub-model corresponding to the dynamic range dimension is: the proportion of overexposed highlights is detected to be 18%, and the image quality dimension score is c1; the output result of the sub-model corresponding to the noise dimension is: the estimated noise level is 1.2 times the expected value of the database mean, and the image quality dimension score is c2; the output result of the sub-model corresponding to the color naturalness dimension is: the color mean of the skin color area is 4.3, and the image quality dimension score is c3. Then, the image quality dimension scores of the above sub-models are weighted and calculated, and the weighted total score is a1×c1+a2×c2+a3×c3. Then, the percentage position is calculated based on the overall data of the database, and normalization is performed to obtain the final target image quality score. The repair suggestion is: It is recommended to enable HDR mode and use a strong noise reduction model to reduce noise in the image.

[0125] Optionally, the target image quality assessment result may be a visual report including a heat map or the like, so that the user can intuitively understand the image quality.

[0126] According to the image quality assessment method based on multi-sub-model collaboration provided by the embodiment of the present application, a dynamic gating network fuses the image feature vector, the context feature vector and the object preference vector to form a fused feature vector, and the model weights of multiple sub-models are calculated based on the fused feature vector, so that the image quality assessment focus can be automatically adjusted according to different image content perception situations and different object preferences, and the sub-model is activated on demand, thereby avoiding static assessment bias and improving the physical interpretability of sub-model selection. It can achieve fast and automatic image quality assessment, improve the efficiency of image quality assessment, and reduce manual intervention and assessment time; through the sub-model division of labor design, each image quality dimension can be evaluated separately , effectively avoiding multi-task conflicts, enabling image quality assessment to cover both low-level image quality degradation and high-level perceptual attributes, and setting up a collaborative mechanism between sub-models to further improve the accuracy and efficiency of image quality assessment; this solution provides a more comprehensive image quality analysis and description, and obtains the final image quality assessment result by integrating the output results of the sub-models, thereby enhancing the image understanding and description capabilities, and realizing subjective and objective fusion evaluation. It can not only provide quantitative image quality scores, but also provide qualitative image quality description information and repair suggestions, improve interpretability, and solve the defects of existing image quality assessment methods in evaluation dimension coverage, scene adaptability, and interpretability.

[0127] Figure 3 FIG shows a structural block diagram of an image quality assessment device based on multi-sub-model collaboration according to an embodiment of the present application. Figure 3 As shown, the device includes: a feature extraction module 310, a gating module 320, a processing module 330 and a result integration module 340.

[0128] The feature extraction module 310 is adapted to perform feature extraction on the image data to be evaluated, and obtain an image feature vector and a context feature vector of the image data.

[0129] The gating module 320 is suitable for: splicing the image feature vector and the context feature vector by a dynamic gating network to form a fused feature vector, and calculating the model weights of multiple sub-models based on the fused feature vector; wherein different sub-models are used to evaluate image quality from different image quality dimensions.

[0130] The processing module 330 is adapted to: select at least one target sub-model from the plurality of sub-models according to the model weight, and activate the at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension.

[0131] The result integration module 340 is adapted to integrate the output results of at least one target sub-model to obtain a target image quality assessment result.

[0132] Furthermore, the feature extraction module 310 is further adapted to: perform calculations based on the image data to be evaluated to obtain basic statistical features of the image data; extract semantic features of the image data through a neural network; utilize an image feature processing network branch to process the basic statistical features and semantic features to obtain an image feature vector of the image data; utilize a context feature processing network branch to process the basic statistical features and semantic features to obtain a context feature vector of the image data.

[0133] Furthermore, the feature extraction module 310 is further adapted to: obtain object interaction data associated with the image data, perform image quality assessment preference analysis on the object interaction data, and obtain an object preference vector; the gating module 320 is further adapted to: splice the image feature vector, the context feature vector, and the object preference vector by a dynamic gating network to form a fused feature vector.

[0134] Furthermore, the gating module 320 is further adapted to process the fused feature vector using a fully connected neural network and a multi-classification activation function to obtain model weights of multiple sub-models.

[0135] Furthermore, the processing module 320 is further adapted to: compare the model weight of each sub-model with an activation weight threshold, and determine the sub-model whose model weight is greater than the activation weight threshold as the target sub-model.

[0136] Furthermore, the processing module 330 is further adapted to: activate at least one target sub-model, input the image feature vector and the context feature vector into the at least one target sub-model, and have the at least one target sub-model perform image quality evaluation from the corresponding image quality dimension based on the image feature vector and the context feature vector to obtain the output result of the at least one target sub-model.

[0137] Furthermore, the processing module 330 is further adapted to: during the image quality assessment process, different target sub-models exchange intermediate feature information through the attention module.

[0138] Furthermore, the processing module 330 is further adapted to: when at least one target sub-model includes a target sub-model at a first level and a target sub-model at a second level, the target sub-model at the first level is preferentially activated to perform image quality evaluation from the corresponding image quality dimension; if the output result of the target sub-model at the first level meets the basic evaluation requirements, the target sub-model at the second level is then activated to perform image quality evaluation from the corresponding image quality dimension.

[0139] Furthermore, the result integration module 340 is further adapted to: perform weighted calculation on the image quality dimension score in the output result of at least one target sub-model according to the model weight of at least one target sub-model to obtain a weighted total score, and normalize the weighted total score to form a target image quality score in the target image quality assessment result.

[0140] The description of each module above refers to the corresponding description in the method embodiment and will not be repeated here.

[0141] According to the image quality assessment device based on multi-sub-model collaboration provided by the embodiment of the present application, a dynamic gating network fuses the image feature vector, the context feature vector and the object preference vector to form a fused feature vector, and the model weights of multiple sub-models are calculated according to the fused feature vector, so that the image quality assessment focus is automatically adjusted according to different image content perception conditions and different object preferences, the sub-model is activated on demand, static assessment bias is avoided, the physical interpretability of sub-model selection is improved, and fast and automatic image quality assessment can be achieved, the image quality assessment efficiency is improved, and manual intervention and assessment time are reduced; through the sub-model division of labor design, each image quality dimension is evaluated separately , effectively avoiding multi-task conflicts, enabling image quality assessment to cover both low-level image quality degradation and high-level perceptual attributes, and setting up a collaborative mechanism between sub-models to further improve the accuracy and efficiency of image quality assessment; this solution provides a more comprehensive image quality analysis and description, and obtains the final image quality assessment result by integrating the output results of the sub-models, thereby enhancing the image understanding and description capabilities, and realizing subjective and objective fusion evaluation. It can not only provide quantitative image quality scores, but also provide qualitative image quality description information and repair suggestions, improve interpretability, and solve the defects of existing image quality assessment methods in evaluation dimension coverage, scene adaptability, and interpretability.

[0142] An embodiment of the present application provides a non-volatile computer storage medium, which stores at least one executable instruction or computer program, which enables a processor to perform operations corresponding to the image quality assessment method based on multi-sub-model collaboration in any of the above-mentioned method embodiments.

[0143] An embodiment of the present application provides a computer program product, which includes at least one executable instruction or computer program, which enables a processor to perform operations corresponding to the image quality assessment method based on multi-sub-model collaboration in any of the above-mentioned method embodiments.

[0144] Figure 4 A schematic structural diagram of a computing device according to an embodiment of the present application is shown. The specific embodiment of the present application does not limit the specific implementation of the computing device.

[0145] like Figure 4 As shown, the computing device may include: a processor 402 , a communications interface 404 , a memory 406 , and a communication bus 408 .

[0146] Processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408. Communication interface 404 is used to communicate with other devices, such as clients or other server network elements. Processor 402 is used to execute program 410, which may specifically perform the relevant steps of the embodiment of the multi-sub-model collaborative image quality assessment method for a computing device.

[0147] Specifically, the program 410 may include program codes, which include computer operation instructions.

[0148] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application. The one or more processors included in the computing device may be processors of the same type, such as one or more CPUs, or may be processors of different types, such as one or more CPUs and one or more ASICs.

[0149] The memory 406 is used to store the program 410. The memory 406 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0150] Program 410 can be specifically used to cause processor 402 to execute the image quality assessment method based on multi-sub-model collaboration in any of the above-mentioned method embodiments. The specific implementation of each step in program 410 can refer to the corresponding descriptions of the corresponding steps and units in the above-mentioned image quality assessment embodiment based on multi-sub-model collaboration, and will not be repeated here. Those skilled in the art will clearly understand that for the convenience and brevity of description, the specific working process of the above-mentioned devices and modules can refer to the corresponding process description in the above-mentioned method embodiment, and will not be repeated here.

[0151] The algorithm and display provided herein are not inherently relevant to any particular computer, virtual system or other device. Various general-purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious that the structure required for constructing this type of system. In addition, the embodiments of the present application are not directed to any specific programming language yet. It should be understood that various programming languages can be utilized to realize the content of the embodiments of the present application described herein, and the description of the specific languages above is for the purpose of disclosing the best mode of implementation of the embodiments of the present application.

[0152] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures, and techniques are not shown in detail so as not to obscure the understanding of this description.

[0153] Similarly, it should be understood that in order to streamline the present disclosure and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present application, the various features of the embodiments of the present application are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting the following intention: that the claimed embodiments of the present application require more features than the features explicitly recited in each claim. More precisely, as reflected in the claims below, inventive aspects lie in less than all the features of the individual embodiments disclosed above. Therefore, the claims following the detailed description are hereby expressly incorporated into the detailed description, with each claim itself serving as a separate embodiment of the embodiments of the present application.

[0154] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and in addition may be divided into multiple submodules or subunits or subcomponents. All features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed herein may be combined in any combination, except that at least some of such features and / or processes or units are mutually exclusive. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0155] Furthermore, those skilled in the art will appreciate that although some embodiments described herein include certain features that are included in other embodiments but not other features, combinations of features from different embodiments are intended to be within the scope of the present invention and to form different embodiments. For example, in the claims below, any of the claimed embodiments may be used in any combination.

[0156] The various component embodiments of the embodiments of the present application can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all of the functions of some or all of the components in the embodiments of the present application. The embodiments of the present application can also be implemented as a device or apparatus program (e.g., computer program and computer program product) for performing a part or all of the methods described herein. Such a program implementing the embodiments of the present application can be stored on a computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0157] It should be noted that the above embodiments illustrate rather than limit the embodiments of the present application, and that a person skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbols placed between brackets should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The embodiments of the present application may be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc. does not indicate any order. These words may be interpreted as names.

Claims

1. A method for image quality assessment based on multi-sub-model collaboration, comprising: Performing feature extraction on the image data to be evaluated to obtain an image feature vector and a context feature vector of the image data; The dynamic gating network concatenates the image feature vector and the context feature vector to form a fused feature vector, and calculates model weights of multiple sub-models based on the fused feature vector; wherein different sub-models are used to evaluate image quality from different image quality dimensions; selecting at least one target sub-model from the plurality of sub-models according to the model weight, and activating the at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension; The output results of the at least one target sub-model are integrated to obtain a target image quality assessment result.

2. The method according to claim 1, wherein extracting features from the image data to be evaluated to obtain an image feature vector and a context feature vector of the image data further comprises: Performing calculations based on the image data to be evaluated to obtain basic statistical features of the image data; extracting semantic features of the image data through a neural network; Using an image feature processing network branch, the basic statistical features and the semantic features are processed to obtain an image feature vector of the image data; The basic statistical features and the semantic features are processed using a context feature processing network branch to obtain a context feature vector of the image data.

3. The method according to claim 1, before the dynamic gating network concatenates the image feature vector and the context feature vector to form a fused feature vector, the method further comprises: Acquiring object interaction data associated with the image data, performing image quality assessment preference analysis on the object interaction data, and obtaining an object preference vector; The dynamic gating network concatenates the image feature vector and the context feature vector to form a fused feature vector as follows: The dynamic gating network concatenates the image feature vector, the context feature vector, and the object preference vector to form the fused feature vector.

4. The method according to claim 1, wherein the step of calculating the model weights of the plurality of sub-models according to the fused feature vector further comprises: The fused feature vector is processed using a fully connected neural network and a multi-classification activation function to obtain model weights of multiple sub-models.

5. The method according to claim 1, wherein selecting at least one target sub-model from a plurality of sub-models according to the model weights further comprises: The model weight of each sub-model is compared with an activation weight threshold, and the sub-model whose model weight is greater than the activation weight threshold is determined as the target sub-model.

6. The method according to any one of claims 1 to 5, wherein activating the at least one target sub-model to perform image quality assessment from a corresponding image quality dimension further comprises: Activate the at least one target sub-model, input the image feature vector and the context feature vector into the at least one target sub-model, and have the at least one target sub-model perform image quality evaluation from the corresponding image quality dimension based on the image feature vector and the context feature vector to obtain the output result of the at least one target sub-model.

7. The method according to claim 6, wherein the at least one target sub-model performs image quality evaluation from corresponding image quality dimensions based on the image feature vector and the context feature vector, further comprising: During the image quality assessment process, different target sub-models exchange intermediate feature information through the attention module.

8. The method according to any one of claims 1 to 7, wherein activating the at least one target sub-model to perform image quality assessment from a corresponding image quality dimension further comprises: When the at least one target sub-model includes a target sub-model at the first level and a target sub-model at the second level, the target sub-model at the first level is preferentially activated to perform image quality evaluation from the corresponding image quality dimension. If the output result of the target sub-model at the first level meets the basic evaluation requirements, the target sub-model at the second level is activated to perform image quality evaluation from the corresponding image quality dimension.

9. The method according to any one of claims 1 to 8, wherein integrating the output results of the at least one target sub-model to obtain a target image quality assessment result further comprises: According to the model weight of the at least one target sub-model, the image quality dimension score in the output result of the at least one target sub-model is weightedly calculated to obtain a weighted total score, and the weighted total score is normalized to form the target image quality score in the target image quality assessment result.

10. A device for image quality assessment based on multi-sub-model collaboration, comprising: A feature extraction module, adapted to extract features from the image data to be evaluated, and obtain an image feature vector and a context feature vector of the image data; a gating module adapted to concatenate the image feature vector and the context feature vector using a dynamic gating network to form a fused feature vector, and to calculate model weights of multiple sub-models based on the fused feature vector; wherein different sub-models are used to evaluate image quality from different image quality dimensions; a processing module, adapted to select at least one target sub-model from the plurality of sub-models according to the model weights, and activate the at least one target sub-model to perform image quality evaluation from a corresponding image quality dimension; The result integration module is adapted to integrate the output results of the at least one target sub-model to obtain a target image quality assessment result.

11. A computing device comprising: A processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the image quality assessment method based on multi-sub-model collaboration as described in any one of claims 1 to 9.

12. A computer storage medium, wherein at least one executable instruction is stored in the storage medium, and wherein the executable instruction enables a processor to execute operations corresponding to the image quality assessment method based on multi-submodel collaboration as described in any one of claims 1 to 9.

13. A computer program product, comprising at least one executable instruction, wherein the executable instruction enables a processor to execute operations corresponding to the image quality assessment method based on multi-sub-model collaboration according to any one of claims 1 to 9.

Images