Hierarchical Dictionary Learning and Sparse Coding for Static Signature Verification

Elias N. Zois, Marianna Papagiannopoulou, Dimitrios Tsourounis, George Economou; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018, pp. 432-442

An assortment of review papers as well as newly quoted literature clearly indicates that the modeling and subsequent verification of static or offline signatures still poses an active field of scientific interest. Usually, the most important link in the chain of designing signature verification systems (SV's) is the feature extraction one, in which any static signature image is projected and measured onto a feature space. Feature extraction methods are divided in two main categories. The first one, inspired by different computer vision applications, includes handcrafted features. That is features manually engineered by scientists to be optimal for certain type of information extraction-summarization from signature images. Typical examples of this kind include global-local and/or grid-texture oriented features. The second feature category addresses signature modeling and verification with the use of dedicated features, usually learned directly from raw signature image data. Typical representatives include Deep Learning (DL) as well as Bag of Visual Words (BoW) or Histogram of Templates (HOT). Quite recently also sparse representation (SR) methods which include dictionary learning and coding have been introduced for signature modeling and verification with promising results. In this paper, we propose an extension of the classic SR conceptual framework by introducing the idea of embedding the atoms of a dictionary in a directed tree. For the purpose of this work, this is demonstrated by employing an l0 tree-structured sparse regularization norm which has been found to be useful in in a number of cases. We examine the efficiency of the proposed method by conducting experiments with two popular datasets namely the CEDAR and MCYT-75. The verification results of the proposed method are considered at least comparable with those provided with other state-of-the-art approaches.