Study On Sparse Representation Based Off-line Signature Verification

With the improvement of the society, signature is gradually becoming to be a kind ofbiological characteristics widely used for identification. The signature written by hands hasa broad perspective in applications, such as electronic-business, electronic-bank, militaryaffairs, security, office automation and communication. So it is significant for us to deal withthe off-line signature verification research.The research about the off-line signature verification at home and abroad has more than10years, the most mature of which is the support vector machines (SVMs) based method.However, once the SVMs is employed, a specific SVM model must be trained for a signer,and since a model is matched up with a certain amount trained sample, you have to retrainmodel when samples are enlarged. Therefore, it leads to higher complexity and time-consuming as well as weak expansibility of the process.So, we propose a novel classification method based on Sparse Representation (SR) todeal with the problems mentioned above. In this method, the class-specific dictionary can besimply established by using the registered samples (examples). With these registeredsamples, a test signature can be sparsely represented. In order to determine the signature istrue or false, we propose the residual. The lower the residual is, the more realistic thesignature tends to be. In contrast with the conventional classifiers used in signatureverification, such as SVMs, our proposed methods are pretty simple and fast. Moreover,because no training stage is needed and the dictionary can be easily expanded by additionalsamples, this method gains strong expansibility.In order to capture the intra-class variability information more effectively, we designan intra-class variant SR based classifier. What makes this method different from abovemethod is replacing the characteristic of signature with difference value between the featureand its mean value during establishing the dictionary. We also propose a fusion sparse representation to research the comprehension effect of the SR and intra-class SR. Theexperiment result indicates, the accuracy can be highly improved by using these methods.Lastly, we conduct a lot of experiments to demonstrate the effectiveness of theproposed methods compared with the SVMs based verification, in which, we set up rigorousexperiments condition for academic purpose. The images we use in the experiments are allfrom the GPDS960Graysignature database.All methods mentioned above have been contributed to ICIP2014, waiting for theresults.