Currency Detection for Visually Impaired

Accurately identifying currency is a significant challenge for visually impaired individuals, severely impacting their autonomy and financial independence. This paper presents an Android-based application for real-time currency detection designed to assist the visually impaired. The application leverages a TensorFlow Lite model, ModelUnquant, which has been pre-trained for Indian currency classification using lightweight Convolutional Neural Networks (CNN). The model processes images captured through the device camera, which are converted from YUV format to RGB, resized to 224x224 pixels, and normalized for optimal inference. The model outputs a confidence vector corresponding to six currency denominations: ₹10, ₹20, ₹50, ₹100, ₹200, and ₹500. The highest confidence score above a predefined threshold of 60% is used to classify the currency. For efficiency, the model is quantized to reduce computational overhead, enabling near-real-time detection on mobile devices. This solution aims to provide a quick, accessible, and reliable currency recognition experience, enhancing the autonomy of visually impaired individuals in handling currency transactions.