Research On Off-Line Text-Independent Writer Identification

Writer identification is an important biometric identification method, which in public security, justice, archeology, finance and e-commerce has a wide range of applications, and text-independent offline writer identification is one of the most widely used in writer identification. This article focus on research on text-independent offline writer identification algorithm.In this paper, we studied and implemented text-independent writer identification system based on Bag of Words(BoW). We tried to analysis and compare SIFT, SURF, CNN activation feature and LBP. Inspired by Contour-Hinge feature, we proposed ELBP and ESIFT. Experimental results demonstrate ELBP and ESIFT complementory with each other. The final results can be further improved after the fusion of the two features.On feature representation, we first make a comparison between traditional Hard Voting and LLC sparse coding. Meanwhile, we proposed to use LASC in writer identification for the first time. In contrast to HV and LLC, LASC take into account each word around the neighborhood spatial information, which is much better than Hard Voting and LLC method. Moreover, when the dictionary space is larger, these LASC based method does not appear overfitting too early. Meanwhile, we compared the BoW based method with the GMM based method in terms of their advantages and disadvantages as well as their respective performance.Finally, we proposed a multi-feature fusion multi-dictionary based text-independent writer identification system. By fusing the feature of high discrimination, we achieved good performance on the public ICDAR2013 and CVL datasets.