Research On CNN Based No-reference HDR Image Quality Assessment Technology

Image quality assessment,as for a classic subject in the field of image processing,has always been the main focus for lots of researchers.Studies and applications for High Dynamic Range(HDR)image technology has achieved great advances in recent years due to the development and popularization of imaging technology,5G technology and artificial intelligence technology and thus HDR image quality assessment(IQA)gradually becoming hot.The HDR IQA mainly consists of subjective and objective evaluation methods.Considering human visual perception,subjective evaluation is the most ideal way to evaluate HDR image quality.Subjective evaluation methods present high accuracy and reliability,but its operating steps are cumbersome,time-consuming and are of high overall cost and dependence on personnel and hardware equipment resources.Therefore,research for sake of accurate,reliable,simple and efficient objective quality evaluation methods is crucial for HDR image quality assessment.This paper conducts in-depth research on the HDR image quality evaluation algorithm based on the analysis of the advantages,disadvantages and the technical difficulties which current HDR image quality evaluation methods contain.Besides,the paper advances a new HDR image quality evaluation algorithm.The main work and innovations are enumerated as follows:1.A Visual quality perception model: Based on the visual characteristics such as the brightness and contrast sensitivity of the human perception target,a visual quality perception model is designed,which includes error estimation network and perception resistance network.And its training process consists of two stages.The error weight of distorted image is acquired on the first stage through a training process of error estimation network;The optimal perception resistance value is acquired on the second stage through a training process of perception resistance network;Finally,the quality score of the HDR image is calculated by the mixing function.It has been proved by many experiments that ideal results have been achieved.2.A HDR image quality evaluation dataset,‘HDR-combine DB': The paper expands the capacity of the dataset on the basis of several existing public HDR datasets so that the training data of the convolutional neural network are increased.The paper also conducts the subjective evaluation on images that lack subjective evaluation data in the new dataset,and uses INLSA algorithm to align quality score values of different datasets.3.The CNN-based no-reference HDR image quality assessment algorithm,‘HDRIQACNN': The algorithm successfully combines the visual perception model with a introduced human visual attention mechanism.The introduced mechanism can detect the salient regions of the image which can further serve as training data for the neural network.The purpose for introduction of the mechanism is to simulate the scenario where humans will spontaneously ignore information that is not relevant to image quality.The conducted experiments proves that the quality score predicted by the proposed algorithm is highly consistent with the subjective evaluation score,which is corresponded with the subjective perception of human eyes for the quality of HDR images.