Image Quality Assessment Based On Multi-feature Fusion

In today’s life,Images have become the main way of information expression and communication.However,images are easily disturbed during the process of acquisition,transmission and reproduction,resulting in image distortion,which will directly affect the extraction of effective information by humans.Therefore,how to accurately measure the image quality has become a research hotspot in the image field.Subjective quality assessment is the most reliable method to evaluate image quality with high accuracy,but its process is cumbersome,expensive,and time-consuming,which makes it difficult to integrate into practical applications such as real-time quality monitoring and prediction.Therefore,developing objective image quality assessment methods has become very important,with the aim of designing algorithms or models that are consistent with the Human Visual System(HVS).This work focuses on the following two main points:(1)A full-reference image quality assessment method based on human visual perception characteristics is proposed.In order to overcome the problems that the features are difficult to select and the pooling strategy is difficult to explain,this paper fully considers the perceptual characteristics of HVS,extracts a variety of visual features and uses machine learning method for regression prediction.First,based on the visual attention mechanism and contrast sensitivity characteristics of HVS,the visual saliency feature and spatial frequency feature of the image are extracted respectively.In addition,the chromaticity feature and gradient feature of the image are extracted to describe the color information and structural changes of the image respectively.Then extract the mean,standard deviation and entropy features of the similarity map of the reference image and the distorted image to form feature vector,and finally use random forest to establish a regression model.In the four authoritative databases,after a variety of experimental verifications,it proves the effectiveness and robustness of the method in this paper,which is more consistent with the human visual system.(2)A no-reference image quality assessment model based on dual-domain feature fusion is proposed.Aiming at the problem that it is difficult to accurately predict the image quality caused by insufficient image feature extraction.This paper analyzes and extracts perceptual features in both spatial and frequency domains to simulate complex visual perception mechanism.First,in the spatial domain,naturalness,color and entropy features are extracted.Second,in the frequency domain,the normalized first-digit distribution of wavelet coefficients and discrete cosine transform(DCT)coefficients,as well as the energy subband ratio and frequency variation features of DCT blocks are extracted.The features extracted in the dual domains are then fused to form quality-aware feature vector.Finally,input the feature vectors and subjective scores into the Gaussian process regression tool to train and build a regression model.Experimental results on three databases show that compared with mainstream methods,the proposed method shows good predictive performance and generalization ability,and maintains high consistency with the human visual system.