Offline Signature Identification and Verification Using Capsule Network

In offline signature identification and verification systems, hand -crafted feature extraction methods, such as local binary patterns, have recently been set aside for automatic feature extraction methods such as convolutional neural networks (CNN). Although these CNN-based algorithms often obtain satisfying results, they require either many samples to find the best data representations or pre-trained network weights. To obviate the necessity of many samples as well as pre-trained weights, Capsule Network has recently claimed to achieve the best data representation using only a limited amount of data. This network not only obtains many variations of limited input samples via affine transformations in the algorithm but also uses hierarchical layers to select the most informative features without losing the exact informational position of the others. It is from this point of view that this paper first aims to evaluate performances of Capsule Network and the CNN-based equivalent model for the signature identification task. This evaluation is done under two lower resolutions than is usual to understand whether texture patterns are still staying as informative as they usually are for both algorithms. While Capsule Network achieves 98,8% and 98,6 % accuracies for 64×64 and 32×32 input resolutions, respectively, CNN obtains 55,4% and 54,7% accuracies. The second aim of the paper is to generalize the capability of Capsule Network concerning the verification task. Through this evaluation, the capability of Capsule Network is shown to obtain better feature extraction and classification results compared to the CNN-based equivalent model for the verification task as well as the identification task.