Application of Super Resolution for Optical Character Recognition in Low Quality Images

Machine learning has become a very popular method in various branches of industry and has been successfully applied to a number of practical tasks. Optical character recognition, which is one of the most challenging task in computer vision, has made significant progress due to machine learning applications. Modern OCR systems can provide a high-accuracy predictions both for scanned documents and real-scene images. Despite such power, such models are still suffering from low-quality images, especially, in extreme cases of compressed images. A traditional approach to overcoming such deep learning model weakness is to extend the dataset in such a way as to cover such distortions. However, it requires model retraining and can not guarantee the same accuracy on the previous dataset. We tackle this issue from another perspective. In this paper, we discover how a super-resolution preprocessing step could help the OCR model to recognize images itself. Based on our custom synthetic dataset, we built a super-resolution system. We also performed a careful analysis of how loss functions should be used for text images. Finally, we showed that a custom-trained super-resolution system shows much better results in terms of restored image quality and text recognition accuracy.