Off-line Writer Verification based on Small Segments of Handwritten Text and Convolutional Neural Networks

This paper proposes a new method for writer identification based on small fragments of handwritten text randomly obtained from a paragraph. The main contribution of this work is to show that small fragments carry enough biometric information for writer identification. A second contribution is the creation of 2 repositories of images of handwritten text from 50 writers. The first one is made up of 4 text paragraphs of 64 words in high resolution per writer. The second one contains more than 700 thousand fragments of text per writer. Experiments were conducted with different Convolutional Neural Networks, considering the VGG-16, VGG19, InceptionV3, ResNet-50 and MobileNetV2m models. 2 classification schemes were implemented. First, the classification of individual fragments and, second, the classification of groups of fragments. The best results were obtained using groups of fragments, achieving accuracy of 96% on the identification of a text with the same content and of 87% on the identification of the writer considering a text with different content.