Research On Methods Of No Reference Image Quality Assessment Based On Deep Learning

As the carrier of information transmission,the image will be damaged during the transmission and the image quality will be reduced.Therefore,a good image quality assessment(IQA)method is needed to evaluate the image quality in order to detect the degradation of the image quality in time and take effective measures to deal with it.As a type of IQA method,no reference image quality assessment(NRIQA)method has better advantages,it does not need to rely on the reference image and only needs to evaluate the quality of distorted image based on the characteristics of the distorted image.This undoubtedly brings great convenience to image quality assessment,and also improves the practicality of IQA methods.No reference image quality assessment is designed to independently assess image quality without the need for original reference images.Since the early NRIQA methods based on distortion types are only suitable for specific distortion situations and lack practicality,the proposed general NRIQA method becomes challenging.This article starts with deep learning and uses deep learning for NRIQA.The main research contents of this article are as follows:(1)Starting from the two domains of space and gradient at the same time,a multi-branch convolutional neural network(CNN)is constructed to evaluate the quality of no reference images.The proposed method includes two parts: a score module and a weight module.The score module has two CNN branches.These two branches learn features from the spatial domain and the gradient domain,respectively.Residual blocks are used in the spatial domain branch to construct a feature pyramid network to achieve the representation of multi-scale features.The gradient domain branch uses a common CNN structure.The position feature is introduced in the weight module to measure the contribution of local image patches to improve the prediction performance of the final image quality.Through experiments on LIVE,CSIQ and TID2013 databases,it is proved that this method has a high consistency with subjective perception of human eye.(2)The restoration network is used to recover the distorted image to assist the IQA,which makes the image quality evaluation more consistent with the human visual characteristics.The method includes two sub-networks,an image restoration sub-network and an image quality assessment sub-network.The image restoration sub-network is based on U-Net,and its purpose is to restore the quality of the input distorted image and generate an image that has recovered the quality of the input.The image quality assessment sub-network uses the image restoration sub-network to derive a horizontal feature pyramid.This network relies on the image restoration sub-network to improve the accuracy of image quality prediction.The SROCC and PLCC on the LIVE database are 0.971 and 0.969 respectively,which fully proves the effectiveness of the method.