Off-line Signature Recognition Technology Research

Due to the popularity of computer and Internet, communications among people become more and more frequent. Despite of the convenience, it comes with some security problems. Traditional methods for verification and recognition fail to satisfy the society demands. Fortunately, development of the techniques for recognition and verification of identification based on biometric features provides a more convenient and more reliable solution. As a kind of behavioral feature, signature has advantages of easy acquirement and flexiability in authority. It turns out one of the most popular features in identity recognition and verification.Signature images being the objects, preprocessing, feature extraction, feature selection and recognition are studied in depth in this paper.An effective method for preprocessing of signature image is proposed when preprocessing, starting at the representation. Four information carries are achieved by different operations such as filtering, bianization, thinning, normalization and so on. These carries (normalized gray signature, normalized binary signature, normalized framework and gray framework) give an effective representation which paves the road for a perfect feature extraction.During feature extraction and selection phase, four features are abstracted, including pseudo dynamic feature, shape, texture and a kind of local feature along with the performance comparison. Experimental results show that pseudo dynamic features, direction of grey skeleton, a geometrical feature, Zernike moment center moment based shape descriptor and grid information provide us with a better description, receiving a higher recognition rate. However, other features could not adapt off-line signature recognition well, including topological features, co-occurrence matrix and geometrical moment.We adopt here two ideas in the study of technique for recognition. Firstly, combing KNN and D-S theory, a KNN recognition method based on model and evidence theory is presented. For the lazy learning existed in origin KNN, this new technique takes the influence of samples to be recognized and distance among k neighbors on classification in consideration. At the same time, it provides a solution for the unrecognizable cases. Secondly, a new method for off-line signature recongnition based on neural networks and evidence theory is proposed to overcome the difficulty in training, where the neural network with a lager scale is decomposed into several smaller networks. Evidence theory is employed here to fuse the results of smaller networks, which receives a high accuracy.