Research On Image Quality Assessment Methods Based On Feature Fusion

Image quality assessment(IQA)algorithms aims to evaluate the visual quality of images through quantitative methods.It plays an important role in the production,transmission,processing and consumption of images and videos.It can optimize images and help people get a better viewing experience.Therefore,it has high research value.In different application scenarios,limited by algorithm defects and few datasets,the existing methods is still insufficient in the quality assessment of authentic distorted images under No-Reference(NR)conditions and synthetic distorted images under Full-Reference(FR)conditions.This paper mainly focuses on these two types of problems.The specific contents are as follows:In view of the complex distortions in the image,including large-scale overexposure,artifacts,blurring,and imperceptible local noise,as well as the differences in quality perception caused by different distortions and different content to observers,this paper proposes a NR-IQA algorithm based on multi-scale feature adaptive fusing.Our method extracts multiscale features and fuses them adaptively based on distortion,avoiding ignoring distortions of different granularities while focusing on key distortions that affect image quality.In addition,this algorithm designs a content perception module,which interacts position-related distortion features with deep semantic features and completes quality scoring,fully considering the important role of image semantic information in the process of human visual perception.Experimental results show that our method achieves higher accuracy and generalization than other methods on authentic distorted IQA datasets.For the quality assessment of synthetic distorted images,existing FRIQA methods cannot directly supplement the key semantic and texture information lost in distorted images,and show insufficient generalization and accuracy when the number of IQA datasets is limited.Aiming at the above problems,this paper proposes a FR-IQA algorithm based on mask fusing.This algorithm uses reference image blocks to supplement the lost information in the distorted image through mask fusing.At the same time,the method reduces the dependence on the complete reference image,and improves the robustness of the model by using a specially designed sampling strategy as a data enhancement method.Experimental results demonstrate that the proposed method achieves state-of-the-art performance on multiple synthetic IQA distorted datasets,while exhibiting good generalization.