Research On Quality Assessment Of Tone-mapped Video And Tone-mapped Image

High dynamic range(HDR)visual content is an accurate representation of a real scene,giving a good visual experience.HDR images can only be backward-compatible with conventional display after the compression of dynamic range which is called tone-mapping.Regardless of what kind of tone mapping method is used,certain distortions are inevitably introduced,resulting in a degradation in the quality of tone-mapped video(TMV)and tone-mapped images(TMI).To obtain high quality TMI and TMV,an effective evaluation tool is essential.Therefore,based on the distortion characteristics generated by tone mapping,this paper studies the quality evaluation methods of TMI and TMV.The specific research contents are as follows:(1)Considering the characteristics of TMI itself,this paper proposes a TMI quality evaluation method based on aesthetic features.The aesthetic attributes of the image are extracted from five aspects: color fidelity,color temperature distribution,contrast,darkness and naturalness.The transform domain features are extracted in the gradient,tensor,Gaussian and Laplacian domain.Then the support vector regression is used to yield quality assessment model.The model was tested on the ESPL-LIVE HDR database.The PLCC reached 0.791 and the SROCC is 0.766,which was consistent with the human eye perception results.(2)A reduced-reference TMV quality evaluation method is proposed.Based on the assumption that the high-quality TMI should retain the details of the original HDR image as much as possible in the highlighted area and the low-light area,the video frame brightness is divided in a percentage manner.The degree of information retention is calculated in each highlighted and low-light area.Due to that color distortion has a negligible effect on the quality of TMI,it measures global color information,and also take other aesthetic attributes into account.In the time domain,the SIFT flow algorithm which is robust to the changing of luminance is used for motion estimation.Then extract motion features on motion vectors.In addition,for the scintillation artifacts in the TMV that significantly affect the perceived quality,the extracted flicker features are detected by analyzing the curve of perceived luminance along the time.Finally,random forest is used to predict video quality.The experiment was verified on the TMVD2017 database.The PLCC reached 0.7871 and the SROCC is 0.7265.(3)A no-reference TMV quality evaluation method is proposed through the in-depth analyze of the distortion in the TMV without the original HDR video.An adaptive threshold segmentation method is used to divide the brightness of video frames.In view of that the color-sensitive cone cells are replaced by light-sensitive rod cells in dark vision,the brightness information entropy and naturalness is extracted in both dark and bright areas while extracting color sensitivity in moderately bright areas.From the perspective of computer vision,the surface singularity of the video is captured by Surfacelet decomposition,then the transform coefficients are further analyzed to extract the spatial-temporal motion characteristics.Finally,the above features and other quality-related features are combined to form a feature vector,and the video quality is predicted by random forest.The experimental results on the TMVD2017 database show that the method is in good agreement with the subjective perception results,with PLCC reaching 0.8052 and SROCC being 0.7541.