Research On Segmentation Of Semantic Objects

With the rapid development of internet and big data,mining information of interesting for users from large amount of data is an urgent and challenging research task.It not only depends on information retrieval and data mining technology,but also benefits to the model construction of analysis and understanding tasks.As the foundation for multimedia analysis and understanding,semantic segmentation for image and video can not only extract region of interest,but also alleviate the influence of noises.Due to the ambiguity of multimedia data,semantic segmentation is still a very challenging task.In our thesis,we focus on the research of image and video semantic segmentation.In terms of the difficulties of semantic segmentation in clutter background,this thesis aims at how to mine semantic information and judge object region.To this end,we study a series of mid-level semantic segmentation and human interaction-based high-level object segmentation researches.The details are listed as follows.1.We first research on the problem of middle-level semantic segmentation based on color feature and contour information,i.e.,superpixel segmentation.In this research,we propose a minimum spanning tree based superpixel model to evaluate the region consistency in clutter background.This method could not only generate regular and accurate superpixels in linear time without iterations,but also extract regions with object local topology structure.2.Because of the difficulty of special object extraction in complex background,we focus on the research of object segmentation by high-level information of object detection.A minimum spanning tree based segmentation model is proposed by considering the relationship between local parts.The segmentation model is robust to the inaccurate object detection and noise background.3.In terms of the complexity appearance model of multiple objects,we research on interactive multiple object segmentation.This work proposes a multiple object segmentation model based on formal concept analysis(FCA)and convex shape prior,which could efficiently and effectively segment multiple object without any updating for different object regions.The introduced FCA and convex shape prior convert the multiple label optimization to a binary optimization,which significantly reduce the computational cost.The proposed non-iteration interactive multiple object segmentation overcomes the limitation of complex optimization model and time-consuming computational processing.4.Since it is difficult to obtain accurate object prior due to motion blur,illumination change and clutter background existing in video sequences,we focus on the research of video complexity awareness and part-based segmentation propagation model.The former analyzes the background of video frames and introduces a series of indications of video background complexity,which could efficiently capture the object segmentation in simple frames.The latter constructs a part-based segmentation and propagation model,which could overcome the noise interference of the similarity between foreground and background.The proposed model could handle with the inaccurate object prior in complex background.5.Because existing methods lack available guidance for the video object annotation,we propose a global consistency aware query strategy based video object segmentation method.This method not only introduces global consistency for candidate region to evaluate the annotation likelihood,but also uses region sampling to approximate the segmentation label change of the annotation region space.It indicates a frame who has the region with the most likely annotation probability.Meanwhile,the segmentation results have a large gain by annotating the indication frame.This work overcomes the difficulty of multiple object annotation in complex video sequences.6.The existing methods extract object regions based on the assumption of simple background,regardless of object missing and variant scale object,which is not suitable for suddenly scene change,different shot cuts,uncertain motion pattern and object scale variant.We propose a scale aware object detection based segmentation model for long video sequences.We research on object segmentation for complex long video sequence and propose a variant scale object detection model.The experiments demonstrate that this method can deal with the object segmentation with scale variant in long video sequence.