Research On Blind Image Quality Assessment Based On Visual Perception Interaction

The aim of this study is to address the issue of image distortion in visual media applications,ensuring users obtain a satisfactory visual experience.By developing reliable image quality assessment algorithms,this research aims to promote the advancement of image processing algorithms.Based on this background,the focus of this study lies in the development of accurate and efficient objective image quality assessment(IQA)methods to meet these application requirements.This study begins by exploring the perceptual interaction mechanisms of the human visual system(HVS)and derives three important conclusions.Firstly,there exists consistency between image content and distortion in perception,meaning that images with similar content exhibit similar visual quality under the same distortion conditions.Secondly,image content and distortion exhibit perceptual interactions.That is,different image contents can have varying visibility impacts on the same distortion.Finally,subjective quality labels have limitations,as they primarily reflect the degree of image quality degradation and do not provide information regarding image content and distortion types.Based on the aforementioned conclusions,this study proposes multiple image quality assessment models and algorithms.Chapter 3 presents a content similarity-guided label assignment method,which achieves automated image quality annotation and large-scale pretraining dataset generation.The motivation behind this method is that images with similar content exhibit comparable visual quality under the same distortions.Subsequently,this chapter introduces a local and global feature-dependent networks that evaluates image quality by considering long-term dependency relationships between local and global features.To address the issue of subjective quality scores failing to reflect image content and distortion types,Chapter 4 proposes a quality assessment method by retrieving the instances based on the similarity of content and distortion.The motivation behind this method is that images with similar content and distortion in the feature space based on image content and distortion tend to exhibit similar perceptual quality.To achieve this objective,the proposed method retrieves similar instances from an IQA database through two classification modules: the semantic-based classification(SC)module and the distortion-based classification(DC)module.The test image is first compared to pristine images with similar content using the SC module.Then,the DC module queries similar instances within the distorted images corresponding to each retrieved pristine image.Finally,the predicted quality score is obtained by aggregating the subjective quality scores of these similar instances.Considering that retrieval instance-based model is primarily designed for evaluating images with synthetic distortions,Chapter 5 aims to develop a general-purpose quality assessment model.This chapter presents a method for assessing image quality by leveraging the interaction between content and distortion.Firstly,a dual-stream structured network is designed,consisting of a content perception module(CPM),a distortion perception module(DPM),and a visual interaction module(VIM).The CPM and DPM are used to extract semantic features and distortion features from images,respectively,while the VIM is employed to fuse these two types of features and capture their interaction.To simplify the model parameters,this chapter further presents a single-stream structured network,which combines knowledge distillation and distortion recognition tasks and models the interaction between semantic and distortion information in an end-to-end manner.This study comprehensively evaluates and analyzes the proposed models and methods on multiple image quality assessment datasets.The effectiveness of the proposed methods is validated through various evaluations,including single dataset testing,single distortion testing,and cross-dataset testing.Specifically,the content similarity-guided quality annotation method effectively alleviates the limitation of dataset size,effectively improving the model’s generalization capability.Moreover,the model utilizing content and distortion retrieval diminishes prediction bias through the construction of nearest neighbor relationships within the feature space.Finally,the content and distortion-based visual interaction model enhances the accuracy of quality evaluation by considering the mutual impact of content and distortion features.The insights and research achievements of this study are of significant importance in advancing the field of image quality assessment,addressing practical problems,and promoting academic research.