Multi-task learning for pre-processing of printed Devanagari document images with hyper-parameter optimization of the deep architecture using Taguchi method

An excellent text recognition system requires document images to be finely pre-processed. Several conventional image processing techniques have already been implemented to pre-process Devanagari document images by handcrafting features. In contrast with these methods, a deep learning process can be performed that learns the features automatically. In this paper, we have proposed a transfer learning (TL)-based multi-task deep learning (MTL) architecture for pre-processing of Devanagari document images. The MTL approach allows us to pre-process an input image for three pre-processing tasks, viz. binarization, shirorekha removal, and noise reduction, simultaneously. On the other hand, TL helps to transfer the already learned features from a pre-trained network to the existing one and copes with the problem of dataset scarcity. For each branch of the proposed TL-MTL architecture, we have implemented a convolutional encoder-decoder model. Further, the proposed architecture is optimized using Taguchi's optimization method with different network's hyper-parameters as the control factors. The results are then compared to those from the conventional pre-processing methods that are widely used on document images. The comparative results show that the proposed optimized architecture outdoes the traditional image processing methods and has an excellent performance on the dataset of Devanagari document images.

Automated summary

The study introduces a transfer learning-based multi-task deep learning architecture for preprocessing Devanagari document images, which can simultaneously handle three preprocessing tasks and outperforms traditional image processing methods, demonstrating excellent performance in experiments.