Research On Blind Image Quality Assessment Method Combining Visual Perception Characteristics And Deep Learnin

With the development of electronic information technology,digital images as a carrier of information are widely used to transmit information in people’s daily life.However,distortions may occur in images during the process of transmission,compression,storage or camera shooting,resulting in the degradation of image quality.Therefore,how to establish an efficient and accurate objective image quality assessment method to automatically predict image quality scores has become an important issue in the field of computer vision.With the development of deep learning,more and more methods have applied convolutional neural networks to the field of image quality assessment(IQA),which achieved satisfied prediction accuracy.However,distorted images contain rich image content information and image distortion information.It is difficult to train an end-to-end model to extract abundant features that reflect the image quality degradation process.To solve this problem,this work analyzes and models the image quality assessment problem by simulating the human visual system and extracting quality-aware features to improve the prediction accuracy.This work focuses on the following two main points:(1)Existing IQA methods aim to extract distortion features by training deep learning models to predict quality scores.However,images suffer from various distortions.Training a single model is typically hard to handle distortion variation problems.For example,models trained on synthetic distortion datasets are hard to generalize satisfactory accuracy on authentic distortion datasets.To solve the problem,a cycle-consistent adversarial network-based IQA(Cycle IQA)method is proposed to model the image quality degradation process.The proposed method consists of a generative adversarial network(GAN)-based quality perception network and a quality regression network.The GAN-based quality perception network,which consists of a REF2 DIS module and a DIS2 REF module,is designed to simulate the process of distortion information introduced to images in both forward and reverse directions.Inspired by internal generative mechanism of the human brain,the quality regression network extracts the learned hierarchical restoration features from the generator of the DIS2 REF module to learn the relationship between features and quality scores.Experimental results show that the proposed method achieves the satisfied quality prediction performance,especially on the cross-dataset performance.(2)Since current blind image quality assessment methods are not effective in predicting quality scores of both synthetically and authentically distorted images.To solve the problem,a Content Perception and Distortion Inference Network(CPDINet)for IQA is proposed,which divides the IQA task into content perception and distortion inference processes.Since humans try to understand image content before perceiving quality scores,a content feature extractor is designed to explore content information in an image to deal with the content variation problem.To handle the distortion diversity problem,a distortion feature extractor is proposed to capture distortion features in images.Because extracted content features and distortion ones have different characteristics,this thesis proposes adaptive fusion module to fuse multi-scale content features and distortion ones as guidance to selectively enhance important features based on calculated weight scores.Experimental results show that the proposed method can effectively predict quality scores for both synthetically and authentically distorted images.