Research On Key Technologies Of Stereoscopic Saliency Detection

Saliency object detection is a common reserach hotspot in the field of video/image processing and computer vision.It could extract the target area that attracts human visual attention with using modeling analysis for various information in complex scene.The stereo image/video provides the viewer with a more realistic scene space experience and a more intense visual impact.So far,scholars have proposed a relatively mature computational model for the saliency detection of 2-Dvision.However,there is no complete and mature computational framework for the saliency detection of stereo image/video at home and abroad,and the importance of depth features in stereo vision has not been fully explored.Based on the above research status,this dissertation mainly establishes the saliency object detection model for stereo image/video sequence.The main work contents are as follows:First,we propose an RGB-D image saliency detection model with depth credibility factor,which fully considered the contribution of depth information to stereoscopic saliency detection.By calculating the credibility factor of depth map,we can judge whether the depth map can accurately reflect the scene distance or whether the depth map has the scene distortion and blur.When the depth feature can truly reflect the information of three-dimensional space scene,this thesis proposes a saliency detection model combining depth contrast and depth compactness analysis,and proposes a rough background filtering method in this model to suppress the interference caused by the background region with a high depth value.The computational complexity is reduced by calculating salient targets only by depth features.On the contrary,when there is scene blur or distortion in depth map,this thesis proposes an effective calculation method of multilevel ranking saliency based on background priori in combination with color features.At the same time,depth map is fused as the weight of spatial features to calculate the final salient object of RGB-D image.The experimental results on the NJU-2000 and RGBD-1000 databases show that our algorithm achieves 89%and 86%accuracy,and 86%and 80%recall rate respectively.The performance is better than the recent saliency detection algorithm.Based on the saliency detection of stereo image,a prediction model of stereo video sequence is proposed.The model fully considers the spatio-temporal consistency of stereo video sequence and analyzes the motion continuity of foreground objects between adjacent frames.Firstly,spatio-temporal gradient field was constructed by integrating spatial gradient information and motion gradient information,and the foreground possibility based on spatial position prior was obtained as the basis for calculating horizontal geodetic significance of video frame sequence.Then,the scene depth movement information is calculated,and the influence of visual discomfort caused by movement mutation on the saliency intensity is analyzed.Finally,the spatial saliency map and the spatio-temporal consistent saliency map are fused adaptively to obtain the saliency object of the three-dimensional video sequence.The experiment results show that the stereo video saliency model proposed in our work has a good detection effect,and it has achieved 78%accuracy and 72%recall rate,which is better than the recent comparison algorithm,and it can achieve target detection in complex scenes.