Towards generic region segmentation for image/video analysis: An integrated perceptual grouping approach using Generic-Edge-Token-graph

Region segmentation is an important task for many applications in image and video analysis. The process of region segmentation involves partitioning an image into perceptually coherent regions. It serves to simplify image representation from massive individual pixels into constituent compact regions, which form meaningful and efficient representation elements for image content analysis or higher-level object recognition. In recent years, as image and video sources have begun to proliferate, demand has grown for improved region segmentation solutions with low time expense that support both image and video tasks such as video/image indexing and retrieval, object recognition, real-time motion detection/tracking for surveillance, and shot segmentation for video annotation. Despite recent progress, region segmentation still remains a challenging task with respect to robustness, computational cost and solution generality.;A study is provided to demonstrate how the system performs when applied to a video analysis task. The task involves detecting moving objects from video streams. In this application, GET graph is extended into motion GET (MGET) graph with motion attributes added, whereas each moving object is described by a set of perceptual closures in an MGET graph casing GET based semantics hierarchy. Static image region segmentation experiments are also provided for system evaluation. The GRSPG system demonstrates noteworthy potential for supporting various applications that require robust and real-time region segmentation.;This thesis presents an unsupervised Generic Region Segmentation based on perceptual grouping (GRSPG). The system provides precise segmentation results, and produces accurate semantic descriptions for segmented regions while maintaining low computational cost. The strategy is to utilize both edge and region information in the segmentation process by perceptually selecting boundary segments and grouping them into regions based on Generic Edge Token (GET) graph. GETs are perceptually distinguishable linear or curve segments which can be extracted selectively by an edge tracker without processing ail pixels in an image. The system has the following characteristics when compared with the existing region segmentation methods. (1) Instead of computing the entire set of pixels in an image, GRSPG first converts an image into a higher-level feature map, i.e. a GET graph, on the fly. The GET graph can code image content by incorporating global structure and local feature properties of GETs into graph representation. (2) A fast region closure contour algorithm is applied to the GET graph by perceptually grouping adjacent GETs into GET closures (i.e. region contours). The grouping process is controlled by a GET graph search heuristic. (3) The region segmentation approach is robust to both noise and texture, which are handled separately from region contour detection. Noise and texture are measured based on GET's scale, GET graph structure, statistical distribution properties in the GET graph, and region attributes. The output of the GRSPG system contains comprehensive information on segmented regions including size, shape and internal properties for each region.