Writer Identification and Retrieval Using a Convolutional Neural Network

In this paper a novel method for writer identification and retrieval is presented. Writer identification is the process of finding the author of a specific document by comparing it to documents in a database where writers are known, whereas retrieval is the task of finding similar handwritings or all documents of a specific writer. The method presented is using Convolutional Neural Networks (CNN) to generate a feature vector for each writer, which is then compared with the precalculated feature vectors stored in the database. For the generation of this vector the CNN is trained on a database with known writers and after training the classification layer is cut off and the output of the second last fully connected layer is used as feature vector. For the identification a nearest neighbor classification is used. The evaluation is performed on the ICDAR2013 Competition on Writer Identification, ICDAR 2011 Writer Identification Contest, and the CVL-Database datasets. Experiments show, that this novel approach achieves better results to previously presented writer identification approaches.