Research On No-reference Image Quality Assessment Based On Convolutional Neural Network

With the rapid development of information technology and the continuous improvement of people's pursuit of the quality of life,digital images are playing an increasingly important role in people's life.However,in the process of acquisition,compression storage and transmission of digital images,they often suffer from different degrees and types of distortion,resulting in a reduction in the quality of digital images to a certain extent and causing some problems for people's practical applications.The technology of image quality assessment(IQA)is used to evaluate the quality of images and measure the impact of distortion on the quality of images.An efficient and accurate algorithm of the image quality assessment plays a very important role in practical applications and is an important basis for ensuring the user's visual perception experience.Image quality assessment is mainly divided into two categories:subjective IQA and objective IQA,but subjective IQA algorithms are time-consuming,labor-intensive,and have poor practicality.In recent years,objective IQA algorithms have been favored by scientific researchers because they can achieve automatic,fast,efficient and accurate quality evaluation.Because the reference images can not be obtained or are difficult to obtain in most real scenes,the algorithms of the no-reference image quality assessment(NR-IQA)have become the focus of research because they can be free of the limitation of the information with the reference images.Therefore,it is of great research significance and practical application value to develop an efficient and accurate NR-IQA algorithm.People are end users of digital images.In IQA,due to the influence of individual subjective perception differences,the quality scores given by different observers for the same image are often different.Based on the uncertainty of individual subjective perception differences,this paper proposes a variety of effective NR-IQA algorithms by introducing relative quality labels,confidence scores,and multi-task learning into the deep learning framework.(1)A comparative image quality assessment algorithm based on CPNet(Compare-Network)is proposed.This algorithm is an opinion unaware type algorithm that simulates the human visual system to process the relative quality between images,and uses the form of image combination to solve the number limitation of data.The proposed relative quality labels and relative quality order labels have broader application scenarios than absolute quality score labels,and are more convenient and accurate in obtaining.CPNet does not require the reference image,so it is also a NR-IQA algorithm in a sense.The experiments on the LIVE public database prove that CPNet has better accuracy than other algorithms,and its performance is stable.(2)An effective NR-IQA algorithm based on confidence scores is proposed.From the view of statistics,this algorithm explores the uncertainty in IQA due to the perceptual differences of subjective individuals.Based on this,a two-element confidence score is designed to implicitly measure the image quality.The main idea of the algorithm is to represent the visual quality of the image with a two-element confidence score vector.Different from the existing algorithms of directly mapping an image to a quality score,the CSNet(Confidence Score Network)learns the two-element mapping from the image to the two-element confidence score vector.A large number of experimental results from multiple public databases prove that the proposed CSNet is an effective NR-IQA algorithm.(3)A NR-IQA algorithm based on multi-task learning is proposed.Specifically,the algorithm is divided into two sub-tasks: a regression task from images to subjective opinion scores and a classification task from images to image quality levels.Hard parameters are shared in a network model to learn two sub-tasks simultaneously.In the classification task,considering the individual perception differences in the subjective IQA algorithms,the image quality score is mapped to the corresponding level.In addition,a contact module is added between the two sub-tasks to correct the deviation between the two sub-tasks and avoid extreme situations.A large number of experimental results prove the effectiveness and generalization of our method.