Large-scale offline signature recognition via deep neural networks and feature embedding

Calik N., Kurban O. C. , Yilmaz A. R. , YILDIRIM T., Durak Ata L.

NEUROCOMPUTING, vol.359, pp.1-14, 2019 (Peer-Reviewed Journal) identifier identifier

  • Publication Type: Article / Article
  • Volume: 359
  • Publication Date: 2019
  • Doi Number: 10.1016/j.neucom.2019.03.027
  • Journal Name: NEUROCOMPUTING
  • Journal Indexes: Science Citation Index Expanded, Scopus
  • Page Numbers: pp.1-14
  • Keywords: Large-scale dataset, Signature recognition, Convolutional neural network, Batch normalization, Deep learning, VERIFICATION, TRANSFORM, MODEL


Although there have been several developments in offline signature recognition, there is still no much focus on the recognition problem using a small sample size for the training. In many studies, 10 or more signatures are used for training phase, which is mostly intractable in practice. In this study we propose a new convolutional neural network (CNN) structure named Large-Scale Signature Network (LS2Net) with batch normalization to deal with the large-scale training problem. Moreover, we present, Class Center based Classifier (C3) algorithm, which relies on 1-Nearest Neighbor (1-NN) classification task by using the class-centers of the feature embeddings obtained from fully-connected layers. In addition to these, by replacing the activation function rectifier linear unit (ReLU) with leaky ReLU, we create a new network structure called LS2Net _ v2. 96k signatures obtained from 4k signers of GPDS-4000 dataset are used. For a realistic comparison, MCYT and CEDAR are chosen besides the GPDS dataset. The proposed networks are compared with Visual Geometry Group (VGG)-[S, M, 16], which are the frequently used networks in the literature. The networks are tested with two splitting ratios as 25% train - 75% test and 50% train - 50% test per signers. For each ratio, five train and test subsets are randomly generated. Performance metrics are obtained by averaging the results of these five subsets. LS2Net achieved 96.41% and 98.30% accuracy performance for the 25%-75% ratio in MCYT and CEDAR. Moreover, LS2Net_v2 achieves best results by getting 96.91% accuracy score for 25%-75% ratio for GPDS-4000. Batch normalization and C3 algorithm contribute to the performance significantly. (C) 2019 Elsevier B.V. All rights reserved.