Word Spotting in Historical Arabic Documents Using Deep Learning

Nowadays, access to Arabic historical documents has become easier and faster due to the availability of digital copies of manuscripts. However, dealing with them manually remains difficult and costly when indexing, searching, or analyzing data. This paper proposes a deep learning-based word spotting method to predict the word labels. A comparative study was performed by considering several Convolutional Neural Network (CNN) architectures and transfer learning with pre-trained models. Also, two different labels are used; the pyramidal histogram of characters (PHOC) representation and the binary representation. Consequently, two CNN architectures were trained to predict PHOC representation: PHOCNet and CNN with PHOC. In contrast, the rest of the CNN models are trained to predict binary numbers for annotation text labels. To evaluate the models' performance, the Ibn-Sina dataset was used. The best mAP was obtained with the pre-trained model VGG-16, which is equal to 95%; it outperforms the state-of-the-art models.