Bangla Braille Character Recognition with YOLOv8 & Faster R-CNN

Introduction

Bangla is a first language for hundreds of millions of people, yet visually impaired learners still face a steep barrier: most reading and writing tools are built for English, not their mother tongue.

Braille gives touch-based access to text through raised dot patterns, but reading Bangla in Braille is hard, the script is rich, with many characters and subtle differences that are tricky even for people to tell apart. Software that can read these patterns automatically helps close that gap.

This project builds a system that finds and recognises Bangla characters in Braille patterns and turns them into readable text, a practical step toward giving visually impaired students better access to learning in their own language.

Problem Statement

Reading Bangla Braille automatically is harder than it looks:

1. Characters can look almost identical: differing by a single dot or stroke.
2. There is very little ready-made: labelled data to learn from.
3. Dots have to be found: lined up: and grouped into characters.
4. Reading full words and sentences is much harder than reading single letters.
5. Everyday phone photos are messier than clean scanned pages.

Objectives

1. Build a system that detects Bangla characters inside Braille pattern images.
2. Compare a fast: real-time approach against a slower, more precise one.
3. Train on a purpose-built collection of Bangla Braille images.
4. Judge how well characters are found and located.
5. Understand what it will take to move from single letters to full words.

Dataset

The work is built around a purpose-built collection of Bangla Braille images. Depending on the client, this can be data generated from Braille tools, licensed or private material, images from other trusted sources, or a fully custom set captured for the project. The images are labelled, resized, and lightly varied (rotation, shift, brightness, and flips) so the system copes with different layouts and lighting.

Methodology

The project labels Braille images character by character, prepares them consistently, and trains two detectors on the same data: YOLOv8 for fast, real-time scanning and Faster R-CNN for more precise character placement.

The novelty is in pairing these two approaches for a real accessibility tool: a fast detector for live, on-the-go scanning and a more precise one for a careful, accurate readout. This sets up a practical path from simply spotting letters toward reading connected words, the part that matters most for genuine day-to-day use.