Data augmentation for signature images in online verification systems

Abstract

One of the main problems of designing a handwritten signature online verification system is a small number of signatures committed by the user for training. To solve this problem, ways of expanding dataset size based on existing authenticated signatures might be proposed. The research proposes a new technique for generating dynamic signatures based on the original sample. The resulting sample simulates real signature forms and letter-style characteristics. Artificially created genuine and fake samples based on the author’s and intruder’s signatures are used to train the classifier, which can improve the accuracy of training on the original sample of a small size. Handwritten signature data augmentation methods were investigated with the aim of further development in more efficient handwritten verification algorithm.