Quaternion Wavelet Transform Theory And It's Application In Image Quality Assessment

Image is an important carrier for human to obtain information,and it undertakes the important task of transmitting information.In the process of compression,transmission and decompression,a high-quality image may suffer information loss,which causes image distortion.Image quality assessment methods are divided into subjective assessment method and objective assessment method.Subjective assessment method requires human eyes to observe images directly and draw conclusions.Objective assessment method requires computer to assess image quality by algorithmic model.Because the number of images in real life is huge and the distortion situation is complex,human subjective assessment behavior is vulnerable to psychological,physiological,environmental and other factors,so it is difficult to maintain fairness of measurement indicators.Therefore,it is an important task to study objective image quality assessment methods.This paper studies the full-reference image quality assessment of multiply distortion image and no-reference stereo image quality assessment.The main contents are as follows:(1)Quaternion wavelet transform based full-reference image quality assessment for multiply distorted images.Firstly,the method uses quaternion wavelet transform to decompose the reference image and the distorted image simultaneously,and obtains the amplitude and phase of different scales as image features.Then the similarity of features is calculated.Combining the dependence of amplitude and phase with the attention of human eyes to high amplitude regions,the similarity measure is adjusted to fit the human visual characteristics.Finally,the quality score of the whole image is obtained.Experiments show that the SROCC and PLCC indices of this method in multiple distortion image database surpass the traditional full-reference measurement.(2)No-reference stereo image quality assessment based on quaternion wavelet transform.Stereo image is different from 2D image,it has left view and right view,and the two views in the real world are very similar.This method measures the similarity between left view and right view,and uses human visual characteristics to obtain cyclopean image.Then the left view,right view and cyclopean image are normalized by mean subtracted contrast brightness.The normalized coefficients are fitted with generalized Gauss distribution to extract the statistical features of natural scenes,and the features of down-sampled images are extracted again.The samples are trained with support vector regression model to predict image quality.This method is consistent with the subjective perception of human eyes.(3)No-reference stereo image quality assessment based on multi-perception and feature combination.Disparity maps,uncertainty maps,saliency maps and fusion maps of stereo images are extracted,and multi-valued perceptual information is integrated into stereo imagequality assessment,and their statistical characteristics are calculated.Then multi-scale analysis of images is carried out by using quaternion wavelet transform,and statistical features of natural scenes in the quaternion wavelet domain and spatial domain of stereo image pairs are extracted to enrich the dimension of features.Finally,eigenvectors are formed,and support vector regression model training is input to obtain the prediction scores of stereo images.Experiments show that the objective evaluation index of LIVE 3D image database in this paper surpasses the popular stereo image quality evaluation method,and has subjective consistency with human subjective evaluation.