Image Representation And Recognition Based On Complex Network Theory

The rapid development of computer and information technology makes the image recognition technologies have been applied more and more widely. As the basis of the image processing, the image representation has been playing an increasingly vital role in computer vision and image recognition. In recent years, complex network theory is becoming attractive for many scholars. Benefited from the statistical features, the image features based on the complex network have the advantages of favorable stability and strong anti-noise ability. In view of this, a novel image representation and recognition method based on complex network model is proposed in this paper, mainly including the following contents:an image representing based on complex network model; two different kinds of evolutionary methods based on Minimum Spanning Tree decomposition and K-Nearest Neighbor respectively and image feature extraction under dynamic evolution; image representation method combining traditional grey histogram and complex network statistical characteristics, i.e. the vertices weighted complex network model approach, implementing a novel image statistical feature extraction and recognition method.The main contributions and novelties in this paper are as follows:(1) For the structure characteristics usually become instable in traditional graph, an image representation and recognition method based on complex network model is proposed in this paper. An initial complex network is constructed in which nodes correspond to the key points of an image. A novel dynamic evolution process is devised for the initial complex network model based on the minimum spanning tree decomposition to generate a series of sub-networks. The features of the sub-networks in different evolution stages are extracted to finally achieve image structural information extraction. Related experimental results demonstrate that the proposed method outperforms the traditional edge weight threshold evolution method and can describe the structure of images more effectively.(2) In order to describe the image structure more accurately, the direction information is considered in the undirected graph. Therefore, a directed complex network representation model is proposed in this paper. Firstly, we extract key points for an image and construct an initial complex network in which nodes correspond to the key points. Then, a novel dynamic evolution method called K-Nearest Neighbor evolution for the initial complex network model is devised to form a chain of directed sub-networks. At last, we propose a characteristics descriptor by extracting the features of the different stages to achieve image feature extracting and recognition. Related experimental results demonstrate that the method under directed complex network model can achieve better recognition effect.(3) The traditional histogram method lost the spatial position information of the pixel points. Therefore, a method based on complex network nodes attribute (grey value) evolution is proposed in this paper, i.e. the image structural features extraction method based on the vertices weighted complex network model. It combines image content characteristics and structural features, making the traditional histogram feature become one-dimensional in our feature vector. Firstly, an initial complex network is established by representing the pixels as the nodes in the network. Then the dynamic evolution of the complex network is conducted based on the vertices attributes to generate a series of vertices weighted attribute sub-networks. When extracting topological features of the sub-networks, we added the number of the sub-network vertices to constitute a feature vector, and finally achieve image recognition. Experimental results demonstrated that such a discriminative feature extraction framework can achieve superior results in classification compared to conventional well-established image recognition approaches.