Patch-based offline signature verification using one-class hierarchical deep learning

Automatic processing of offline signature verification (in general) can be considered as a low-cost solution to problems in biometrics in comparison with other solutions (e. g. fingerprint, face verification, etc.). This study aims to present a novel writer-dependent approach to verifying an individual's signature through offline image patches of their handwriting. The proposed approach is based on hierarchical one-class convolutional neural network for learning only genuine signatures with different feature levels. Since forgeries are not available for each user enrolled in a real application scenario, this study considers signature verification as a one-class problem. In addition, to achieve a clear structure in image, designing hierarchical network architecture based on the coarse-to-fine principle can lead to more precise results. With lower-level features, the network presents a higher visual quality at the boundary area revealing similarities between genuine signatures, while higher-level features can discriminate the quality of the pen strokes to predict forgeries from genuine signatures. The presented system was tested on two Persian databases (PHBC and UTSig) as well as two Latin databases (MCYT-75 and CEDAR). The results of the analyses produced by this method were generally better and more exact in terms of the four signature databases compared with the present state-of-the-art results.