Malaysian Banknote Recognition App for the Visually Impaired Using Deep Learning

Banknotes are used worldwide, enabling people to exchange products and services. However, visually impaired and blind (VIB) individuals may face difficulties handling banknotes, especially faded or worn banknotes. Therefore, this research proposed a real-time banknote recognition system using Single Shot Detector (SSD) Mobilenetv2 algorithm. The SSD Mobilenetv2 algorithm is efficient, fast, and suitable for low-end devices, making it a suitable technique for this research. A dataset of six Malaysian banknote images (RM1,5,10,20,50, and 100) was augmented and pre-processed was used in this research. Several experiments were conducted to optimize the model's performance using Python, TensorFlow, and Keras. The optimized model achieved 77% accuracy with an average speed suitable for mobile device implementation, using an input size of
$320\times 320$
, 80:20 data split, 32 batch size, and 30,000 training steps. The application was able to detect the notes under various real-time settings like illumination, occlusion, or even half-folded notes, with an audible output voice. Subsequently, further development is necessary to enhance the accessibility and user-friendliness of the application. Increasing the dataset's size and variation of banknotes would help improve the accuracy, and adding vibration alerts when a note is detected would also be beneficial. These improvements empower visually impaired individuals, enhancing their independence and interaction with the world.