Off-line Signature Verification Based On Shape Context

Biometrics is essentially characterized by the use of physiological or behavioral characteristics of people to achieve personal identification. Signature is one of the behavioral characteristics, as its convenience, reliability and strong privacy, it's widely accepted that it has widespread application in electronic-business, military affairs, communication, electronic-bank, office automation and security. Currently,online signature verification technology is gradually mature. Due to a lack of available information of offline signature, it has great potential for development. In the future, there are significant theoretical meaning and practical value on studying offline handwritten signature verification by using computer.Offline signature verification is a difficult task in pattern recognition. According to the requirement of the practical and high accuracy of the products, signature images being the objects, preprocessing, feature extraction and fusion system are studied in this paper. Our research work focus on the following:Firstly, in order to verify the validity and practicability of the proposed method, this paper establishes a Chinese handwritten signature image library by our own lab, the images are preprocessed by denoising smoothing, normalization, binarization and thinning.Secondly, one of the problems of offline signature verification is the lack of use of shape descriptor which will classify the signatures on the basis of their shapes. To solve this problem, this paper proposes using shape context as a shape descriptor, it is established a strong shape descriptor which is set based on the point of the contour, as far as possible the complete representation of the whole image information. And because of the two-dimensional invariance of the polar coordinate transformation, the shape matching can be effectively carried out after the image is transformed by scale, rotation and translation. In order to get better results, this paper uses the Euclidean distance to optimize the matching cost, and improves the matching precision. In this dissertation, we have done sufficient experiments on Chinese database and GPDS960 dataset, the results show that our algorithms is comparable to others.Thirdly, to make up for deficiencies single system, the fusion system is composed of two subsystems. The first stage extracted density gradient direction characteristics of signature, it is a kind of pressure characteristics and can reflect the signature of the force direction, then weighted fuzzy classifier were identified; The second stage extracted shape context of signature, which reflects the shape of the signature information, then uses the template matching method to identify. After the cascade of two classifiers, the decision results are fused and compared with the GPDS960 signature database. Experiment shows that the proposed approach were effective to fuse the advantages of two subsystems and to improve verification accuracy.