Visual Saliency Detection For Stereoscopic Image

Saliency prediction is considered to be a key attentional processing which improves learning and survival by making the organisms to focus their limited cognitive resources and perceptual on the most interesting region of the available sensory data.The computational model of saliency prediction is widely used in various fields of computer vision,such as object detection,scene recognition,and robot vision.In recent year,some comprehensive and well-performing models were proposed.However,these models are suitable for 2D content.With the rapid development of 3D imaging technology,an increasing number of applications are emerging for 3D image or video,which increase the requirement of the computational saliency model for 3D content.Compared to the significant progress in 2D saliency research,the work considering the depth factor for stereoscopic saliency analysis is rather limited,because the researchers are still exploring the role of the depth factor in stereoscopic saliency analysis.This thesis explores the roles of the depth factor from the three aspects:how to leverage the stereoscopic saliency detection by the depth factor,how to build the stereoscopic saliency model based on the mechanisms of the human stereoscopic vision,and how to establish the stereoscopic saliency model that conforms to the 3D content according to the different mechanisms of the human stereoscopic vision.From the three aspects,this thesis proposes the three computation models for stereoscopic saliency prediction based on our past and current research outcomes.The contributions of the thesis are as follows:1.A preliminary saliency model for stereoscopic images is proposed.To improve the depth information for stereoscopic saliency analysis,this model exploits the depth information from three aspects.(1)"We extract the low-level features based on the color-depth contrast features in the local and global search range(local-global contrast).(2)To extract the topological structural from the depth map,the surrounding map based on the Boolean map is obtained as a weight value to enhance the local-global contrast features.(3)Based on the saliency probability distribution in the depth information,we employ stereoscopic center prior enhancement to compute the final saliency.2.Stereoscopic visual saliency prediction based on stereo contrast and stereo focus is proposed.The stereo contrast model measures the stereo saliency based on the color/depth contrast and pop-out effect.The stereo focus model describes the degree of focus based on monocular and comfort zone.After obtaining the values of the stereo contrast and stereo focus models in parallel,an enhancement is performed on both values based on clustering.We conduct multi-scale fusion to form the respective maps of two models.Lastly,we use a Bayesian integration scheme to integrate the two maps(the stereo contrast and stereo focus maps)into the stereo saliency map.3.A computational model for stereoscopic 3D visual saliency based on the three mechanisms of the human visual system.The three mechanisms include the pop-out effect,comfort zone,and background effect,which provide useful cues for stereoscopic saliency analysis.Firstly,we analyzed three mechanisms and find that they can explain the most of the phenomenon about the stereoscopic content.Secondly,we build three models for stereoscopic saliency analysis based on these mechanisms,by adding the control function to enhance our models.Lastly,the relationship of three models is not mutually exclusive.One,two,or three models may appear in one image,depending on the content of the stereoscopic image.In order to accurately determine which model the image belongs to,we propose a choosing strategy to find the best combination of three models to process the saliency map.For implement of our approach,we propose a framework based on the multi-feature analysis.These features reflect three aspects:surround region,contrast,and interesting point.All of these models have been verified by the experiments in two eye-tracking databases and outperforms the state-of-the-art saliency models.