WSNet - Convolutional Neural Network-based Word Spotting for Arabic and English Handwritten Documents

This paper proposes a new convolutional neural network architecture to tackle the problem of word spotting in handwritten documents. A Deep learning approach using a novel Convolutional Neural Network is developed for the recognition of the words in historical handwritten documents. This includes a pre-processing step to re-size all the images to a fixed size. These images are then fed to the CNN for training. The proposed network shows promising results for both Arabic and English and both modern and historical documents. Four datasets - IFN/ENIT, Visual Media Lab - Historical Documents (VML-HD), George Washington and IAM datasets - have been used for evaluation. It is observed that the mean average precision for the George Washington dataset is 99.6%, outperforming other state-of-the-art methods. Historical documents in Arabic are known for being complex to work with; this model shows good results for the Arabic datasets, as well. This indicates that the architecture is also able to generalize well to other languages.