Writer identification for historical handwritten documents using a single feature extraction method

Abstract

—With the growth of artificial intelligence techniques the problem of writer identification from historical documents has gained increased interest. It consists on knowing the identity of writers of these documents. This paper introduces our baseline system for writer identification, tested on a large dataset of latin historical manuscripts used in the ICDAR 2019 competition. The proposed system yielded the best results using Scale Invariant Feature Transform (SIFT) as a single feature extraction method, without any preprocessing stage. The system was compared against four teams who participated in the competition with different feature extraction methods: SRS-LBP, SIFT, Pathlet, Hinge, Co-Hinge, QuadHinge, Quill, TCC and oBIFs. An unsupervised learning system was implemented, where a deep Convolutional Neural Network (CNN) was trained using patches extracted from SIFT descriptors. Then the results were encoded using a multi - Vector of Locally Aggregated Descriptors (VLAD) and applied an Exemplar Support Vector Machine (E-SVM) at the end to compare the results. Our system achieved best performance using a single feature extraction method with 91.2% mean Average Precision (mAP) and 97.0% accuracy.