| This thesis describes an object recognition method. It starts from the image segmentation. We propose a Fuzzy-like Reasoning Segmentation (FSR) method. Unlike other segmentation methods, the FSR method considers illumination condition as one of the segmenting factors. Thus, images taken under different illumination condition expect to get more consistent segmentation results. It improves the segmentation performance in terms of human perception.; And we propose a Local-Global (L-G) graph method to represent segmented image. The local graph describes the information associated with a region; the global graph represents the geometrical structure of image and spatial constraints. In contrast to previous object recognition techniques, which require that object is represented by one region, the L-G graph method can recognize object segmented to several regions. The L-G graph is the describer of the target image, including both shape and structure description. Then, object recognition is implemented via L-G graph searching and comparing. Recognition experiments for multi-view image database are described, in which models are built and tested with real world pictures. |
