Spatial Multi-scale Filtering Based Deep Blind Image Quality Assessment

Since the 21st century,with the rapid development of information technology,digital images/videos have rapidly integrated into all aspects of our lives.Image/video has been widely used as one of the most efficient and intuitive description vehicles.However,image/video often brings different degrees of distortion in the process of acquisition,processing and transmission.We usually need to evaluate the degree of distortion.On the other hand,various image enhancement algorithms need to be objectively and automatically evaluated after enhancing distorted images.Therefore,the realization of a fast,accurate,automatic and objective image quality assessment method has important practical significance and theoretical value.There are two main methods for evaluating image quality.One is subjective quality assessment,and the other is objective quality assessment.The former relies on human eyes for evaluation,which is certainly consistent with subjective feelings,but it is time-consuming,costly,and poorly stable.The latter is automatically completed by intelligent algorithms,which are often simple and convenient,stable,low in cost and highly repeatable.However,some existing objective image quality assessment methods may do not maintain a high degree of consistency with the subjective perceptions of the human eyes.This is due in large part to the fact that method do not capture well the important aspects related to human visual system(HVS).Achieving high consistency with subjective perception is highly challenging.To solve this problem,this paper studies the contrast sensitivity function and multi-channel features of Human Visual System(HVS),and proposes a blind evaluation algorithm for image quality that is more in line with human subjective perception.The innovations in this article are as follows:1)In this paper,the visual information received by the human eye is decomposed by spatial multi-scale filtering to form the characteristics of different frequency bands,The influence of these features on the human visual perception system is explored in turn to find out the features that meets the human visual perception and learn the relationship between input data and subjective scores by means of convolutional neural network fitting and statistical feature extraction.2)Analysis of the advantages and disadvantages of the existing image quality assessment neural network model,combined with image patch-based training and global information extraction,not only absorbs the advantages of image patch-based training to prevent overfitting,but also ensures the reliability of results when the model utilize global information to fit subjective scores,also ensures the stability of the output results when the size of test image is not suitable.3)We also found that the filtering levels needed for the best features of different databases are different.We think this phenomenon is related to how much the high and low frequency information of distorted image lost with respect to reference image.Finally,through solid experiments,we showed that the model has better generalization ability and achieved the good performance in classic databases.