Object Extraction Based On Contour Grouping And Shape Matching

Computer vision is intended to make computers own the ability to understand the visual scene like human being. To this end, object detection is an important task. However, a difficult problem which hampers this task is the disturb of clutter objects in natural scenes.This thesis addressed the problem of detecting object based on an edge map. Edge information belongs to the structural information of an image. Edge information always contains the boundary of objects. Yet, it can not be utilized easily for the detection of many cluttered nonsense noise edges. That makes object detection based on edge information much difficult in natural cluttered scenes.Evidence from psychophysics suggests that humans make combined use of multiple levels of image features to detection objects. Based on this finding we analyze the method of utilizing low-, mid- and high-level visual information and construct a computational framework that combines three different levels of visual information which based on bottom-up feature extraction and top-down shape constrain. The research works done by this thesis mainly are:1. We construct an object detection framework which combines low-, mid- and high-levels of visual information. On the low level we detection boundary information in the image. On the mid level we group the previous detected boundary fragments to form a salient contour based on their spatial context relations. On the high level we use shape information to constrain the grouping process and refine the result of contour grouping which lead to the detection of the whole object contour.2. For the detection of low level boundary information, choose a method different with the traditional edge detection methods. During detection boundary information the method considered different kinds of visual cues including intensity, color and texture. The method avoids much noise results and makes the proposed framework very robust to the existence of clutters.3. As to the problem of discontinuity of detected boundary, we enforce grouping process to get the salient contour which is a mid-level visual representation. And we use the salient contour to guide the process of object detection. Moreover, we proposes a edge inhibition algorithm to solve the problem of detection of inner edges which improve the performance of contour grouping.4. For the utilization of shape information, we improve the Contour Segment Networks model and device a object contour extraction algorithm based on deep first graph searching.