Author Identification from Handwritten Characters using Siamese CNN

The NIST Special Database 19 has been studied to solve character recognition tasks. We present a different study on this dataset, where we do author identification using features learned from handwritten characters. Recent studies in computer vision demonstrate that Siamese Convolutional Networks enjoyed successes in image information retrieval tasks. This technique has previously been applied in the identification of people using face image data producing start-of-the-art performance. We apply Siamese convolutional neural networks in author verification based on the handwritten characters. Employing a pairwise-loss approach, we developed a three-layer Convolutional Neural Network, with three fully connected layers, we achieved verification accuracy of 80% on average with unseen test data.

I. Introduction

In this paper, we present a system for the identification of a writer using handwriting text. Identification is based on learning how each author uniquely writes an alphabet or a digit. To distinguish between authors, a Euclidean distance of and embedding is learned based on how each character is written by different authors using a pair of deep convolutional networks. The pair of networks, called Siamese networks, share the same configuration and weights. The learned embedding and the distance produced by the Siamese network indicates a measure of similarity between characters written by the same author and measure of dissimilarity for characters written by different authors. For a unique character written by the same author several times, the network learns to decrease the distance between characters embeddings. A unique character written by different authors, the network learns to increase the distance between character embeddings. Siamese neural networks models developed by Schroff et al. 2015 [1] for individual identification produced state-of-the-art performance using face images. Siamese convolutional neural networks also produced good results in object tracking among many other applications in computer vision [2].