Art2 Neural Network Clustering-based Certificate Image Retrieval Technology Research

How to use computer technology to judge exactly if the attachment proof materials have been used repeatedly is an urgent problem of the National Science and Technology Awards Accreditation. This paper focuses on the technology of content-based certificate image retrieval, to solve the problem of judging quickly if the materials from a mass of attachment proof materials have been used repeatedly. And, solving the difficult issues of certificate image retrieval, image features retrieval and search efficiency, by focusing on the methods of image features retrieval and the technology of clustering-based hierarchical retrieval.In the study of the methods of image features retrieval, based on the specificity of the certificate Image features, the image is firstly transformed into binary image, then the pseudo-Zernike invariant moment is used to describe the features. On the basis of analyzing the invariance of pseudo-Zernike moment in detail, this paper presents the improved method for pseudo-Zernike moment, making the improved pseudo-Zernike moment keep the rotation invariance, and have the proportion invariance in the true sense. In the aspect of the technology of clustering-based hierarchical retrieval, on the basis of ART2 neural network having the advantage of not destroying the patterns already storied when learning new patterns, ART2 neural network is used to cluster the similarity of the feature moments. It can recognize stored patterns and cluster patterns quickly by adjusting vigilance parameter to harmonize both the stability and flexibility. And, finally, hierarchical retrieval of the certificate images is achieved. In the paper, the drifting ceiling coefficient of ART2 is set to make clustering more stringent, improve the sensitivity of the network to the gradual change model and resolve the problem of insensitivity of traditional ART2 network to the gradual change of input model.Lastly, this paper makes an application modeling using the improved feature extraction method for pseudo-Zernike moment and the improved technology of clustering hierarchical retrieval for ART2 neural network. And, then, it constructs a prototype which achieves a satisfactory experiment results. Experiments show that the clustering performance of the improved ART2 network is better than traditional ART2 network.