Abstract
Optical Braille Recognition methods usually use many designed steps, such as image deskewing, Braille dots detection, Braille cell grids construction and Braille character recognition, which are less robust for complex Braille scenes. This paper proposes an optimal semantic segmentation framework BraUNet to directly detect and recognize Braille characters in the whole original Braille images. BraUNet adds extra auxiliary learning strategy to UNet network, which uses long-range connections of feature maps between encoder and decoder to get more low-level features. And auxiliary learning strategy can combine multi-class Braille characters segmentation with Braille foreground extraction, which can improve the feature learning ability and the Braille segmentation performance. Then morphological post-processing is used on semantic segmentation results to get the final individual Braille character regions. Experimental results show the proposed framework is robust, effective and fast for Braille characters segmentation and recognition on both complex double sided Braille image dataset and handwritten Braille image dataset.