OPTIBRAILLE - assistive device for visually impaired

Objective: The objective of OptiBraille is to develop a low-cost, standalone assistive Braille character recognition system designed and built on a Raspberry Pi 4, enabling visually impaired users to access written Braille in real time. The system aims to provide real-time offline conversion of Braille to text without requiring a network connection or a high-cost device. It also aims to integrate edge AI inference and text-to-speech feedback into a portable assistive device suitable for resource-constrained environments.Methods: OptiBraille captures Braille images using a Pi Camera V2 and preprocesses them through RGB normalisation. A MobileNetV2-based convolutional neural network (CNN) is used to recognise characters A-Z. The model is trained using three-stage progressive fine-tuning and optimised through post-training quantisation to TensorFlow Lite format. A cross-domain generalisation evaluation framework is employed using multi-source training data comprising synthetic, scanned, and real-photograph datasets, with an Angelina-only validation strategy to optimise real-world deployment performance. The V3 model is deployed on Raspberry Pi 4 for real-time, fully offline inference and integrated with text-to-speech feedback. The system achieves 92.55% real-world accuracy on photographically diverse Braille book images, with a quantified cross-domain generalisation gap of 7.45% relative to synthetic test accuracy and a compressed model size of 2.71 MB. The prototype uses a Raspberry Pi 4 and Pi Camera V2.Impact: OptiBraille provides an affordable and portable assistive technology solution that enables visually impaired users to convert Braille into text and receive real-time speech feedback without internet connectivity. Its low-cost hardware, lightweight deep learning model, and offline operation make it suitable for resource-constrained environments. The cross-domain generalisation evaluation demonstrates the system's ability to address the performance challenges associated with real-world Braille images. The architecture can also be ported to custom embedded boards with minimal hardware, supporting wider deployment. Future extensions include multi-line Braille recognition, Grade-2 Braille support, and smartphone integration.