Braille character detection using deep neural networks for an educational robot for visually impaired people

Teaching computer programming to the visually impaired is a difficult task that has sparked a great deal of interest, in part due to its specific demands. Robotics has been one of the strategies adopted to help in this task. One system that uses robotics to teach programming for the visually impaired, called Donnie, has as its key part the need to detect Braille characters in a scaled-down environment. In this paper, we investigate the current state-of-the-art in Braille letter detection based on deep neural networks. For such, we provide a novel public dataset with 2818 labeled images of Braille characters, classified in the letters of the alphabet, and we present a comparison among some recent detection methods. As a result, the proposed Braille letters detection method could be used to assist in teaching programming for blind students using a scaled-down physical environment. The proposal of EVA (Ethylene Vinyl Acetate) pieces with pins to represent Braille letters in this environment is also a contribution.

A Histogram-Based Two-Stage Adaptive Character Segmentation for Transcription of Inter-Point Hindi Braille to Text

Problem statement: The optical Braille character recognition (OBR) system is in substantial need in order to preserve the Braille documents to make them available in future for the large section of visually impaired people. The recognition and transcribing of the double sided Braille document into its corresponding natural text is indeed a challenging task. This difficulty is due to the overlapping of the front side dots (recto) with that of the back side dots (verso) in the inter-point Braille document. In such settings, the habitual method of template matching to distinguish recto and verso dots is unproductive. Approach: A fresh system for double sided Braille dot recognition is proposed, which employs a two-stage highly efficient and an adaptive technique to differentiate the recto and verso dots from an inter-point Braille expending the horizontal and vertical projection profiles along with distance thresholding for Braille character segmentation. Materials: The efficacy of this segmentation technique is demonstrated on a large dataset consisting of Hindi Devanagari Braille documents with varying image resolution and with diverse word patterns. The primary reason for choosing the Hindi Devanagari Braille is that, Hindi is the national language of India and OBR for the Hindi Devanagari Braille is not available. Results: Braille line segmentation accuracy of 100%, word segmentation accuracy of 99.8% and character segmentation accuracy of 99.4% has been accomplished. Conclusion: This effort of OBR development for Hindi Devanagari Braille has been done for the first time. The proposed method is tolerant to merging of Braille dots and presence of half characters.

E-Braille Documents: Novel Method For Error Free Generation

Present manuscript proposes a technique for estimating the angle of rotation for a Braille document image which in turn will be used for aiding its automatic character recognition. The technique is based on maximizing number of null projection of the derived image vector. Results show that any amount of rotation transformed distortion can be nullified and thus leading to proper reading of imprinted Braille character pattern. The proposed methods have been successfully tested on manually written as well as computer generated Braille with rotation distortion.