Unsupervised feature learning for writer identification
Document typeMaster thesis
Rights accessOpen Access
Our work presents a research on unsupervised feature learning methods for writer identification and retrieval. We want to study the impact of deep learning alternatives in this field by proposing methodologies which explore different uses of autoencoder networks. Taking a patch extraction algorithm as a starting point, we aim to obtain characteristics from patches of handwritten documents in an unsupervised way, meaning no label information is used for the task. To prove if the extraction of features is valid for writer identification, the approaches we propose are evaluated and compared with state-of-the-art methods on the ICDAR2013 and ICDAR2017 datasets for writer identification.