Researches On Image Quality Evaluation Methods For Panoramic And Stereo Applications

Image is a major source of information for human perception and machine pattern recognition,and its quality plays a decisive role in the adequacy and accuracy of the information obtained.With the continuous deepening of social informationization,wide-viewing angles,high-resolution panoramic images,and stereoscopic images with a sense of presence,interaction,and immersion experience have become research hotspots in the field of image processing.Panoramic image can make people express information more effectively and understand the real world better,while stereoscopic image can bring rich,lifelike and vivid experience effect to viewers,which has great significance in people's production and life.Therefore,in order to better apply stitched image and stereoscopic image to various fields of society,this paper focuses on the evaluation of the stitched image and stereoscopic image in the following three aspects:(1)A color-corrected stitched image quality assessment method was proposed.Firstly,a stitched image database with five kinds of color differences was established by using the existing color correction algorithm and stitching algorithm.Secondly,to evaluate the quality of stitched images with color difference comprehensively,four features were extracted from the pre-splicing image sequence and the stitched image respectively.Then the four features were combined to establish a relation model between features and quality through support vector regression algorithm,so as to predict the quality of stitched images with color difference.The proposed method was tested on the stitched image database with color difference.The experimental results show that the proposed method can effectively evaluate the quality of stitched images with color difference.(2)A quality assessment method for stitched images with bi-directional matching is proposed.Firstly,the testing and benchmark stitched images undergo bi-directional SIFT-flow matching to establish dense correspondence.Then,color distortion,geometric distortion and structural distortion are measured respectively and fused via support vector regression(SVR)to get the final quality score.A large number of experiments were carried out on the stitched image database with color difference,including 400 testing images.Experiments on the established database demonstrate the superiority of our method.(3)A blind quality assessment method for asymmetric multiply-distortion stereoscopic images based on high order statistics aggregation(HOSA)was proposed.Firstly,the training samples with diversity and universality were selected,which have five views of left,right,cyclopean,sum,and difference image.They also include three kinds of distortion combinations:asymmetric multiply-distortion,symmetric single distortion,and asymmetric single distortion.Then the HOSA features of each view of the training samples to train the global dictionary and test the global quality of each view,and obtain the global quality through SVR algorithm.The local dictionaries were trained to get local quality to complement and improve the overall prediction accuracy of stereoscopic image.Finally,the global quality and the local quality are combined by linear weighting to obtain the final quality.The experimental results show that the proposed method can effectively evaluate the quality of stereoscopic image with various distortion types.