Stereoscopic Visual Saliency And Its Application On Stereoscopic Image Disparity Control

Visual saliency extracts region and object that draw our interest and fixationfrom images. It has been widely used in image segmentation and object recognition.Compared with traditional visual attention model, stereoscopic visual saliency hasbetter performance due to the addition of depth feature. Although stereoscopic imagestechnology are blooming in the last few years, the research focus on generating,transmitting and displaying the stereoscopic content from the angle of the maker andthe subjective perceptions of end users have always been underestimated. Disparitycontrol technology based on visual saliency could overcome the fusion problem ofmulti-view stereoscopic images and makes it possible that the target region or objectcan be rendered at different depth of the auto-stereoscopic display and improve thestereoscopic perceptions of end users. Stereoscopic visual saliency and its applicationon stereoscopic image disparity control are studied deeply and widely in this thesis.In the aspect of stereoscopic visual saliency, this thesis begins with the modelingand evaluation procedure of visual saliency, and then goes deep into the influence ofdepth feature towards visual saliency. The depth contrast and histogram thresholdstereoscopic visual saliency model are proposed based on region contrast saliencymodel. Furthermore, the salient object segmentation ground truth database isestablished manually. The thesis evaluates the proposed stereoscopic visual saliencymodel subjectively and objectively. PR curves illustrates that depth feature canimprove the performance of visual saliency model effectively.In the aspect of vergence control, this thesis derives the vergence controlprinciple for parallel multi-view stereoscopic images. Then the adaptive vergencecontrol is implemented based on Mean Shift image segmentation. At last, thevergence control platform oriented to multi-view stereoscopic images is establishedusing visual saliency model and SIFT operator and natural3D scene could bereproduced vividly when the multi-view images that have been processed by vergencecontrol are rendered on the display.