Shape Representation And Matching By Latent Semantic Structure

Machine system based on visual perception has a wide range of applications, such as: intelligent video analysis, intelligent traffic, scene recognition, battlefield perception, scene matching guidance, remote sensing image analysis, image retrieval, automatic navigation, robot automatic picking, and so on. However, the performances and applications of the machines which have the ability of visual perception are based on the level of object's representation and recognition. The performance of object's representation and recognition has become the technical bottleneck to implement machine's visual perception ability. And at the same time, it is the core problem and key technology to any system with the ability of visual perception. Shape is one of object's most basic sense features, so the research of shape representation and recognition has been the hotspot problem in the field of computer vision.The biggest difficulty to shape representation is shape can be influence easily by the factors of deformation, distortion, and so on. At present, both of the shape contour's global and structural representation methods are depend on constructing a robust description to eliminate the influences of all kinds of factors, but the methods that merely relying on building description can't fundamentally solve this problem. For this, this paper proposes a novel research approach, which is the shape latent semantic structure model (SLSSM). Compared with previous methods, SLSSM no longer relies on constructing robust description to eliminate the influences of all kinds of factors, but firstly allow the large interference factors to influence the shape's local contour, then getting the real semantic relation among the local contours by semantic analysis methods, thus obtaining shape's latent semantic structure.SLSSM consists of these parts: feature detection, shape segmentation, local contour description, building shape words and getting shape semantic. Among them, shape word and shape semantic are this paper proposed new concepts, and at the same time they are the keys and difficulties of SLSSM. The concept of shape word is similar with the document's word , represents a kind of local contours which are similar in form. When shape is influenced by large interference factors, its local contours'differences will be very large, such local contours will be expressed as different shape words, but the semantic relations among them is same. It is meaning that the shape words as the same as the document words, having the semantic ambiguity phenomenon of synonyms. Therefore, this paper obtains the real semantic relations among shape words by semantic analysis methods, thus getting shapes'latent semantic structures.Finally, this paper demonstrates the effectiveness of SLSSM in various common shape databases. These shape databases consist: MPEG-7 CE-Shape-1 shape database, Kimia99 shape database, Kimia 216 shape database, Aslan and Tari's shape database which contains 56 shapes. At the same time, by comparing with shape context algorithm, it further demonstrates the effectiveness of SLSSM model.