AI-Powered Currency Recognition for the Visually Impaired

• Shubham Banwal

• July 2023

A Sustainable ML-Based Currency Recognition System

Problem:

Visually impaired individuals face significant challenges in identifying physical currency, leading to dependency and risk of financial errors in everyday transactions.

Approach:

Designed and developed a real-time computer vision system that uses mobile camera input to detect and classify currency denominations, combined with voice feedback for accessibility.

System Design:

• Designed a CNN-based currency recognition system using TensorFlow to classify currency denominations for visually impaired users.
• Built an image preprocessing pipeline covering resizing, normalization, lighting variation handling, and standardized input preparation.
• Developed a CNN architecture using convolutional, pooling, fully connected layers, ReLU activation, and dropout regularization.
• Trained and evaluated the model on currency images with varied denominations, orientations, and lighting conditions, achieving approximately 95% accuracy.
• Proposed integration with mobile/smart assistive devices for real-time currency recognition and audio-based user feedback.

Key Contributions:

• Built a computer vision pipeline to detect and classify currency using image recognition techniques.
• Developed a real-time inference system enabling instant denomination detection via mobile camera input.
• Integrated text-to-speech functionality to provide clear audio output for user interaction.
• Designed the system with a focus on usability, accessibility, and low-friction interaction.
• Optimized for accuracy and performance under varying lighting and positioning conditions.

Constraints & Tradeoffs:

• Variability in image quality, lighting, and note conditions
• Need for fast, low-latency inference on mobile devices
• Accessibility-first design (minimal interaction, clear output)