Research On Visual Comfort Improvement For 3D Program Production

Nowadays,major 3D display techniques take advantage of two cameras to present a more immersive quality of experience.The disparity information between the two images enables the depth perception of the viewers.However,the accompanying phenomenon of visual comfort reduction and visual fatigue of the viewers has caused widespread concern.In general,we can divide the factor that affects human visual comfort into two categories,one is the error brought in during the program shooting,like the vertical disparity or the synchronization problem of the two cameras,and the other comes from the inherent contradiction between the 3D display techniques and the human vision model,for example the Accommodation-Convergence Conflict induced by excessive disparity,or the binocular fusion difficulties induced by Window(or Edge)Violation.It's an important work for 3D TV industry to look into the varies factors that affect 3D visual comfort and to establish an effective scheme for 3D program visual comfort improvement.In this thesis,firstly,an overall framework for 3D program comfort improvement is proposed,and implemented in the form of a plug-in for Adobe After Effects,a common-used video compositing application.The plug-in leverages the rectification algorithm to eliminate the vertical disparity in the video,and enables synchronization of the two camera by frame shifting in the timeline.Moreover,it can achieve disparity adjustment by disparity shifting and image retargeting,and it also support the auto calculation of the adjusting parameters for the non-key frames.Based on that,a model-driven 3D visual comfort improvement method is proposed,so as to provide a reference for the disparity adjustment of the key-frame images and to achieve better viewing experience.In the first place,a comfort assessment algorithm is proposed for the 3D images with multiple salient regions.Actually,there are already tons of researches published in the field of 3D visual comfort,yet visual factors considering the gazing model and the disparity jump during the image viewing are hardly mentioned.We believe these factors have potential connection with visual comfort experience.To be specific,we first propose five kinds of visual features related to the visual saliency area in the 3D images,which include salient region disparity distribution feature,disparity jump between the salient regions,disparity jump around the salient regions,distribution of the salient regions and widths of the salient regions.We then establish a 3D image database and conducted subjective visual comfort evaluation experiments,so as to further study these visual features.Moreover,Support Vector Regression is used to build the visual comfort assessment model,based on the proposed five kinds of visual features and the traditional overall disparity distribution feature.The result shows that our proposed features can achieve a satisfactory performance gain.Secondly,we leverage the proposed visual comfort assessment model to establish a 3D image comfort improvement method.In general,this method can be divided into two steps,which is 3D visual comfort assessment model training and disparity adjustment.For the first part,we enhance the proposed visual comfort assessment model by filtering the features.The3 D saliency extraction is performed by combining depthmap-segmentation-based foreground extraction and 2D saliency extraction.And for the second part,we perform disparity adjustment in the usual way—disparity shifting.One dimension search is used to find the local optima of the disparity shift values.Meantime,the searching scope is restricted so as to prevent side effects like the aggravation of Accommodation-Convergence Conflict and the aggravation of the DepthCue Conflict.For the objectiveness of the evaluation,model training and disparity adjusting is performed on the IVY LAB 3D image database.Lastly,subjective evaluation experiment is conducted to give a comparison between the images before and after the adjustment,the result of which shows that the performance of this method meets our expectation.