Study On Contour Image Recognition Based On Affinity Propagation Clustering

Image recognition is an application of pattern recognition technology which developed greatly in recent years, an important research area in Artificial Intelligence, also a common branch of computer vision. It carries out its perception and understanding just like human based on modern information processing technology and computer technology. It is a great challenge that how to make computer observe the environment, understand the outside world and even benefit mankind. Shape matching is fundamental for many tasks such as object recognition, matching, registration, and image retrieval. It seems that researchers are more interested in retrieval by shape than by color and texture. The key point of target profile identification is the description of the contour feature. Recognition effect is related to the generality and accuracy of the feature description directly.In order to solve the problem of the complexity and instability of existing contour feature description algorithms, a novel method based on hierarchical feature description of contour image is proposed. From the perspective of human cognition, a few contour feature descriptors that represent global, local and detail characters respectively is proposed. And then these characters are combined into a four dimension vector, which is regard as feature to distinguish objects with each other. The Euclidean distance between feature vectors is used to measure the similarity between different shapes.For feature recognition and matching, the common approach is to compare the similarity between the two images directly, or use the classification or clustering method. Contrary to the classification method, clustering is a unsupervised algorithm and the demand for the division of data between classes is higher. In order to test the effectiveness of the proposed feature descriptor, clustering method is used here to identify the image. Since we have analyzed and compared some classical clustering algorithms, we choose the affinity propagation clustering algorithm here. The experimental results on Kimia-99 database indicate that this hierarchical description is more advantage then the traditional feature description methods. Furthermore, it performs better in shape recognition and image similarity matching in terms of robustness when compared with other contour-based methods in literature.