Research On Shape Description Based On Complex Network

Shape description is an important research area in image processing and pattern recognition, it is widely used in object identification and medical image processing Recently, lots of methods have been proposed by researchers, however, shape images of real objects have rigid deformations and non-rigid deformations, which result in the efficacy losing of a lot of proposed methods. Shape analysis based on complex networks is an important method and it belongs to the boundary based shape description methods. It doesn’t take inner information of shapes into consideration, but considers the positions of boundary points only. Based on the method, shape models based on multi-model complex network and directed complex network are proposed in this paper.(1) For traditional complex network models of shape images, weights of edges between nodes is Euclidean distance of corresponding boundary points, however, Euclidean distance is instability for shape deformations. Different from Euclidean distance, inner distance is stability for shape deformations. In order to solve this problem, multi-model complex network is established which combines Euclidean distance and inner distance of shape boundary points. In the new established model, weights between nodes are determined by Euclidean distance and Inner distance between boundary points. In dynamic evolution stage, based on Euclidean distance and Inner distance two sub-networks are obtained. Then, extract features of sub-networks and describe shapes based on the extracted feature. Experiment result shows that the proposed method is more robust than the single-model complex network model.(2) The dynamic evolution of traditional complex networks is based on threshold value, undirected networks are obtained at each evolution stage. Compared with undirected networks, directed networks contain more structure information including distance information and neighbor information. Based on this, a directed complex network model based on inner distance is proposed. Besides, weights of nodes in the network are inner distance in order to be stability for shape deformations. In dynamic evolution stage we use k-NN evolution and directed sub-networks are obtained. Compared with traditional complex networks, the proposed method contains more information of the shape images, so it is more robust. Experimental results show that the proposed method can perform better than classical shape representation models.