Multilevel Feature Fusion for No-Reference Image Quality Assessment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image quality assessment methods, particularly no-reference algorithms, face challenges in accurately assessing image quality without reference images, especially for GAN-generated images, leading to inefficiencies and inaccuracies.
Innovation Solution
An image assessment method utilizing a multilevel transformation network, fusion network, and fully connected layer to extract and fuse image features, improving prediction accuracy and stability by processing images through a neural network model with twin branch networks and employing joint loss functions for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple image features are extracted and fused through a multilevel transformation network, then image quality assessment accuracy is improved, but model complexity increases
Solution Approach 1:
The model is divided into multiple transformation networks operating at different levels (low-level, medium-level, high-level features). Each network extracts specific types of features independently, and these segmented feature extraction processes are then fused to achieve comprehensive image quality assessment. This segmentation allows the system to manage complexity by breaking down the feature extraction task into manageable, specialized components.
Solution Approach 2:
The patent introduces a multilevel hierarchical dimension to feature extraction, processing images at multiple scales and abstraction levels rather than a single level. This dimensional approach allows the model to capture both fine-grained local features and coarse-grained global features simultaneously, improving assessment accuracy while organizing complexity in a structured hierarchical manner.
2Adaptability or versatility
If a no-reference algorithm is used to assess images without reference images, then applicability to practical tasks is improved, but measurement precision deteriorates
Solution Approach 1:
The patent introduces learned feature representations as intermediaries between the input image and the quality assessment output. These intermediate features, extracted through the multilevel transformation networks, serve as mediators that capture essential image characteristics without requiring direct comparison to reference images. This intermediary representation enables accurate no-reference quality assessment by translating complex image properties into assessable feature vectors.
Solution Approach 2:
The model transforms the image assessment problem from direct quality evaluation to feature space evaluation. By changing the parameters from pixel-level comparisons to learned feature-level comparisons across multiple transformation levels, the system achieves both no-reference capability and high measurement precision through the rich feature representations generated by the hierarchical networks.
Data Source
AI summary
The present disclosure relates to an image assessment method and apparatus, and a device, a storage medium and a program product. The method comprises: acquiring an image to be assessed; and inputting said image to be assessed into an image assessment model, so as to obtain a quality assessment result corresponding to said image to be assessed, wherein the image assessment model comprises: a multilevel transformation network, a fusion network and a fully connected layer; the multilevel transformation network is used for processing said image to be assessed to obtain image features, which are output by each layer of transformation network; the fusion network is used for fusing the image features, which are output by the each layer of transformation network, so as to obtain a fused image feature; and the fully connected layer is used for processing the fused image feature to obtain the quality assessment result.