Experimental Study on Scanning of Degraded Braille Books for Recognition of Dots by Machine Learning

Braille books follow processes of deteriorations different from ordinary printed books because visually impaired people read the sharp dot with the finger’s belly. In frequently read braille books, braille is dirty with aging, collapses, and holes open. Then, visually impaired people cannot read braille books. For braille books manually transcribed, the only method to republish them is to rebuild though the original is often and already disposed of. Now, we read degraded braille books with a scanner and restore them. In this paper, comparative experiments on what types of scanner are appropriate to successfully read degraded braille from old braille books are carried out. The candidates are two flatbed scanners, GT-S 650 and Plustek OpticBook 4800, and a stand scanner, ScanSnap SV600. Each scanner scans three types of braille papers, fresh books, used books and old books. Next, to recognize and identify dots in braille papers, we use two methods, the Haar classifier in OpenCV as image processing function and the deep learning with Google’s TensorFlow. For aiming at improving the recognition rate, we also use scanner options suitable for reading. Experimental results show that the flatbed scanner is suitable for reading.