BrailleSegNet: A novel methodology for Braille dataset generation and character segmentation

Abstract

Recent research in the field of Braille learning has highlighted vital role of accurately segmenting Braille letters from Braille documents to improve accessibility and educational opportunities for visually impaired children. A novel methodology, BrailleSegNet, is proposed for Braille Dataset Generation and Braille character segmentation, structured into six distinct phases: Image Acquisition, Image Preprocessing, Fixed-Sized Square Conversion, Rows Extraction, Zonal Operations, and Braille Character Extraction. The initial phase involves acquiring images from the Braille-TextStory dataset, followed by preprocessing steps including grayscale conversion, binary conversion, Gaussian filtering for noise removal, and image inversion. Subsequently, the method standardizes the varying sizes and shapes of Braille dots into fixed-sized squares, extracts rows containing Braille characters, and performs zonal operations such as vertical dilation, zone identification, full zone conversion, and space zone addition to accurately segment and recognize Braille characters. The final phase extracts Braille characters from the designated full zones. Addressing the scarcity of datasets with appropriate ground truth reflecting real-world Braille document scenarios, a new dataset, Braille-TextStory, was created as part of this work. This dataset includes short stories in English, generated using the Braille-PageMap algorithm for evaluating Braille character segmentation techniques. The Braille-TextStory dataset maps English letters to their corresponding Braille images, accurately placing them on plain pages with proper management of parameters such as letter spacing, word spacing, and line spacing to preserve the integrity and readability of Braille documents. The proposed segmentation methodology was tested using this dataset, demonstrating a high level of effectiveness and accuracy compared to state-of-the-art methods.