Research On No-reference Image Quality Assessment Algorithm Based On Multiple Annotators

With the development of multimedia technology,various electronic devices have gradually entered in daily life.As the receiving terminal of information,we get what we need from the mass of information everyday.Digital image as one of the important media of mass information dissemination has been deeply into every corner of life.However,in the process of image collection,storage,transmission and display,different degrees of noise may be introduced due to differences in hardware or software,so that the image quality perceived by users is poor and it is difficult to meet the user's target requirements.One of the urgent problems in the existing image processing field is how to design an efficient image quality evaluation algorithm to truly apply it to practice.The most reliable and effective quality evaluation is human subjective evaluation.Due to the existence of many influencing factors in the real situation,such as time-consuming,laborious and unstable evaluation results,these factors to some extent hinder the application of human subjective quality evaluation in practice.By exploring the possible solutions to the existing problems in the field of image quality assessment(IQA),this paper focuses on the research of objective no-reference image quality assessment.In order to evaluate the quality of natural images,we design efficient and robust IQA model.The research content of this paper mainly includes the following two aspects:(1)In terms of the problem of insufficient data,an effective method of image data amplification is proposed.Traditional IQA depends on the image database containing a small amount of image content.The high latitude of image space is often in conflict with the uncertainty of finite image test.This imbalance may cause a large deviation of the IQA models.Therefore,we propose a method to augment the existing image training dataset by artificially adding a variety of synthetic distortions.Introducing reasonable and comprehensive distortion types and levels,the large scale available image datasets are constructed,so as to realize the image data amplification on quality and content.At the same time,the existing IQA algorithms are used to generate the quality score of distorted image dataset as the data label,and prepare for the subsequent algorithms design and model construction.(2)Simple and efficient IQA methods are proposed for the related research of opinion-unaware image quality evaluation.Traditional blind image quality assessment(BIQA)algorithms are studied via using image feature extraction and image subjective score.The processes of feature extraction and model prediction are not well integrated.In addition,the existing BIQA methods based on data-driven usually use the network with other targets in computer vision tasks as the foundation of model parameter tuning,or training the network through image block decomposition and quality related data introduction,and then finetune the model with some subjective data.Although these methods have obtained good experimental results,their model are complex and fail to deal with the problems of global and local quality approximation.Therefore,combining with deep learning,we propose two methods of no-reference IQA based on multiple annotators to realize the combination of score regression and ranking learning.The experimental results show that the proposed methods can effectively predict the perceived quality of images.The research contents of this paper include image data amplification,network model construction and quality score prediction,etc.The existing problems in IQA are deeply explored,and different thinking angles and solutions are given to improve the overall performance of the no-reference IQA.The research results of this paper can provided theoretical basis and technical guidance for the related research in no-reference IQA.We expect more new ideas to be put forward in the future.