Representation Learning and Dissimilarity for Writer Identification

Writer identification by using manuscripts became a very important research topic in forensic analysis of documents. That is because the writing can be considered as an identifying characteristic of a person. By analyzing the challenges and the resources available for the research development in this field of investigation, one can find different databases of writers manuscripts. In this work, we aim to evaluate the performance of deep learning techniques in the process of writer identification by using manuscripts. For this, a convolutional neural network (CNN) was developed to address the writer identification task. We have also evaluated the performance of features obtained with CNN when submitted to a support vector machine (SVM) classifier, considering the traditional classification approach. To assert the objectives of this proposal, CVL and BFL databases were used. The experiments were conducted by using a texture generation approach, starting from the original documents. Feature learning was accomplished using CNN on the texture obtained from the manuscripts. Finally, we have also evaluated the impact of the classification using both classification scenarios with or without using the dissimilarity approach. The best results were achieved using the SVM classifier and dissimilarity approach on the features obtained with CNN. The performance demonstrated the robustness of the proposed method with performance similar or superior to that described in the literature.