Abstract:
Computational analysis of historical documents has remained an interesting research area for the pattern classification community. Identifying the documents based on structural similarity among their features is a major challenge that constrains the writer identification. The focus of this study is the identification of writing styles from short historical manuscripts based on the structural similarity in handwriting. The documents under study have limited text hence identifying writer-specific features from these images poses a challenging problem. In our proposed technique, the documents are binarized using a deep learning-based model. Small writing patches are then extracted from the documents and are fed to a Siamese neural network with different models as convolutional base. Unlike the classical framework of recognition where the model is expected to learn class labels, a Siamese network supports learning similarities between samples of the same class and differences between samples coming from different classes. We formulate the task of scribe identification as similarity learning problem and employ a contrastive loss function. Models are trained using positive and negative pairs and evaluated on both gray-scaled and binarized patches in order to compare the performance on the both data sets. Among these investigated base models better results are obtained on gray-scaled data. We achieved an overall accuracy of 63% with customize ConvNet and with pre-trained ConvNets up to 73% accuracy is achieved. Considering the complexity of the problem, the reported performance is indeed quite promising.
