Research On Referenceless Image Quality Evaluation Based On Human Visual Attention Mechanism

With the widespread application of digital imaging systems,the importance of Image Quality Assessment(IQA)is increasing.The subjective evaluation method is cumbersome,costly,time-consuming,and cannot be adjusted in real-time.While the corresponding objective image quality evaluation can avoid the tedious cost and waste of time resources under the subjective method,achieving the goal of adaptively perceiving image quality.Image fidelity is used for most of the current mainstream image quality evaluation algorithms,rather than based on the perception of image quality.Image fidelity is just one of the factors to measure image quality,and it is not directly related to image quality itself.At the same time,mainstream image quality evaluations are mostly based on objective evaluation criteria,which are not suitable for human visual features and deviate from the actual discriminative perception of human visual perception.Unlike traditional image quality evaluation methods,non reference-image quality evaluation standards are explored in the article with two approaches: feature extraction+regression/fitting framework and end-to-end framework,as well as extensive experiments.Firstly,various non reference image quality assessment methods in recent years are reviewed: the subjective and objective quality evaluation methods of images,especially some representative works in Non-Reference quality evaluation,as well as commonly used datasets and evaluation indicators in the IQA field.Two aspects are included in the research:(1)To address the low generalization,an improved Structure Nature Perceived quality Natural Image Quality Evaluator(SNP)SNP-NIQE algorithm based on existing non reference image quality evaluation algorithms is proposed.Firstly,the two-stage cascaded significance region detection is introduced into the original SNP-NIQE algorithm to facilitate the re-determination of clustering centers for images with significance regions;Then,gray fluctuation feature and information entropy feature are extracted in turn and fused.Finally,the objective evaluation value of image quality is predicted with machine learning.Through experimental verification,the algorithm proposed in this paper can accurately reflect the perception effect of human vision on image quality with stronger generalization and fewer parameter quantities.(2)A reference free image quality evaluation framework based on deep attention networks is proposed to address factors such as the lack of effective utilization of human visual characteristics and the insufficient ability of the network to extract features in conventional reference free image quality evaluation methods based on deep learning methods.This framework can effectively quantify human visual features into convolutional neural network layers,and make full use of multi-scale information and channel information in the model to enhance the judgment of high-frequency details.A large number of experimental results show that the algorithm has achieved significant performance in both subjective and objective effects.