Study On An Object Recognition Method Based On Non-symmetry And Anti-packing Model

Object detection and localization is a difficult problem because objects in a category can vary greatly in shape and appearance. The variation arise not only from changes in illumination, occlusion, background clutter and view point, but also due to non-rigid deformations, and intra-class variation in shape and other visual properties among objects in a rich category. How do we deal with the variation, especial the intra-class and pose variability of object? Most of the current researches have focused on modeling object variability, including several kinds of deformable template models, and a variety of part-based, fragment-based models.Two methods have been presented to deal with the variation of object. One method can deal with the whole object variation and the other can deal with the local part variation of the object. Firstly, we propose a Non-symmetry and Anti-packing Object representation Model (NAOM) to represent an object category. The NAOM object model consist of several local parts, we call it sub-patterns. The model codes the global geometry of the generic visual object categories with spatial relations linking the objects' parts. The model is similar to the pictorial structure model, k-fans model, constellation model, and star model described above, but it is more flexible and simpler.Secondly, the descriptors of sub-pattern can deal with the local variation of the object. Shape based information have been selected as a key component of local features. After reviewing existing edge based descriptors, our experiments show that Grids of Histograms of Edge Direction (GHED) descriptors significantly outperform existing feature sets for shape class detection.The proposed NAOM model and GHED feature descriptor are embedded in a three layers hierarchical tree structure. The NAOM model at the middle layer organizes all parts to represent the object with the relative spatial relationship between its parts; and each part is described by GHED descriptors at the bottom. The object detection is a bottom-up processing that consistent with the hierarchical tree structure. The proposed framework can be applied to any object that consists of distinguishable parts arranged in a relatively fixed spatial configuration. Our experiments are performed on images of side views of horses; therefore, this object class will be used as a running example throughout the paper to illustrate the ideas and techniques involved.The main contributes of this paper lie in:Firstly, motivated by the hierarchical object model detection method, we propose a novel three layers hierarchical tree structure for object detection and location. Our hierarchical structure can capture the shape deformation of an object class accurately in the whole global object and the local part of the object.Secondly, we propose a Non-symmetry and Anti-packing Object representation Model (NAOM) that is embedded in the middle level of our three layers hierarchical structure.Thirdly, we propose a contour shape descriptor-Grid of Histogram of Edge Direction (GHED) that can deal with the variation of the local shape of an object effectively.