Research On Salient Object Detection Method In Video And Its Application

Salient object detection is one of the popular research subjects of computer vision.The purpose of salient object detection is to predict the object location which firstly attracts the visual interest when people watch image or videos.Salient object detection can assign limited computing resources to region of interest,but also help people to further understand the mechanism of human visual attention.This dissertation studies the detection of salient objects in video from the viewpoint of improving the accuracy,calculating efficiency and increasing the depth information,and apply the salient object detection to video parsing.Video saliency detection using object proposals.We propose a method that take proposal as mid-level feature to detect salient objects in video.This method can realize background prior more direct and accurate,and solve the problem of highlight the edge of object and incomplete object detection in pixel/superpixel scale.Our method takes proposal as new computing unit so that we redefine traditional saliency model based on individual video frames and sequences.According to salient score in different salient features,we apply the strategy of voting and ranking to proposals and get the initial saliency map.Meanwhile,we use the object edge optimization and spatial and temporal consistency to further refine salient object to get the final saliency map.This method is tested in three public datasets and outperforms other state-of-art methods in precision-recall rate and MAE score.Rapid and effective spatiotemporal saliency detection.We propose a method that take motion information as main salient features to detect salient objects in video.This method can further improve the efficiency of computing to 13 fps,which is close to real time computing,on the base of detecting accuracy and robustness.Traditional salient object detection method takes appearance features as mean discriminate salient features.However,when the background is cluttered,object structure is complex,or the motion object is blurred and deformed,the detection result will be inaccurate and inconsistent in temporal domain.According to the human visual attention mechanism,people are more likely to notice moving object when watching dynamic scenes.Therefore,the low-level motion information in video can offer more discriminate salient features.This method takes main motion vectors as primary salient features to locate salient objects in video sequence,and takes center-surround contrast and uniqueness as secondary salient features to optimize salient object detection.It realizes the spatial and temporal consistency in the framework of multi-features.Learning to detect stereo saliency.We propose a learning method based on support vector method to detect salient objects in stereoscopic video,and put out a dataset including 400 pairs stereo images(BIT400)to publicly test salient object detection methods in stereoscopic scenes.There is horizontal minor difference in left and right scenes,and this difference can be presented as disparity maps by stereo matching.In the salient object detection of stereoscopic scenes,disparity maps offer additional visual deep information to further improve the testing accuracy,but also reflect the pixel matching in left and right scenes to keep the consistency of stereo saliency detection results more effectively.We proposal two new features in our method: monocular salient features,binocular salient features and motion salient features.We feed these features to the training and testing steps in SVM to make full use of the ability of generalization of the learning method,then the optimal test results are obtained.Video paring method guided by visual salient object.We propose a nonparametric video parsing method based on multiple deep features that apply salient object detection to video parsing.By improving the accuracy of pixel dense matching,introducing saliency can improve the final performance of video parsing.Meanwhile,the method also proves that as the preprocessing of computer vision,salient object detection can help to solve the high-level semantic visual problems.This method takes the salient object detection results as weights in dense matching to stress significant regions which needs precise matching.Deep features in different levels can reflect various characters of scenes.So we apply multilayer deep features to parsing method.Fully connected features contains more global and semantic information,which can be used to search similar scenes of test frames in retrieval datasets.Dense matching between retrieval scenes and test scenes is built as a coarse-to-fine matching process using multiple convolution features.The matching result is integrated into a highorder spatial and temporal label transfer model to get pixel-level video paring results.