Developing an AI-Enabled Currency Identification App for The Reserve Bank of India

To ensure reliable offline performance, the app was designed to function without internet connectivity. Since it could not learn in real time, Daffodil trained the AI models specifically for offline usage. A proprietary dataset of over 150,000 images of Mahatma Gandhi Series and Mahatma Gandhi (New) Series banknotes was created, covering denominations of ₹10, ₹20, ₹50, ₹100, ₹200, ₹500, and ₹2,000.

The dataset captured variations in orientation, lighting conditions, partial or half-folded notes, camera types, and backgrounds to maximize recognition accuracy. Using Artificial Intelligence (Optical Character Recognition), these images were fed into machine learning models. To expedite development and maintain precision, transfer learning was employed, leveraging pre-trained ImageNet models. Features learned from ImageNet were applied to the custom banknote dataset, enabling the models to quickly and accurately classify currency notes by denomination. Once trained, the models were converted into mobile-optimized formats to ensure efficient offline execution, with automatic flashlight adjustment for low-light conditions.

Additionally, the development plan incorporated security protocols, such as vulnerability. The application underwent rigorous testing to ensure reliability, security, and performance. Vulnerability Assessment & Penetration Testing (VAPT) and Static Application Security Testing (SAST) were implemented to identify and mitigate potential security and performance vulnerabilities at both code and system levels.

Additionally, the app was tested across multiple real-world scenarios, including varying lighting, half-folded notes, and different device environments, to validate the accuracy and robustness of currency recognition. Cloud services such as AWS S3 for image storage and EC2 for scalable computation were integrated into the development workflow, secured via AWS Identity and Access Management (IAM), to optimize both cost and performance during training and testing phases.

In response to diverse linguistic needs, our solution features voice commands and voice overs in 11 regional languages. When users open the app, they are presented with an intuitive menu displaying language options, each one being clearly announced. Users can effortlessly choose their preferred language using simple verbal instructions.

Let’s consider a native Punjabi speaker using this application. They have the flexibility to opt for Punjabi as their preferred language setting. As they scan a currency note, the app delivers a clear and audible announcement of the denomination value in Punjabi. This inclusive feature is designed to accommodate users from diverse regions, allowing them to seamlessly navigate the app in their chosen language.

To enhance accessibility further, the application includes a greeting audio file that guides visually impaired users through its features and usage instructions. However, integrating this audio file was challenging due to its large size. For maintaining uninterrupted offline functionality, it was imperative to uphold a minimal application size. 

Our team adeptly resolved this issue through meticulous optimization, allowing the application to offer a seamless user experience while maintaining its compact design and essential features.

For people with hearing and visual impairments, the app has a predefined number of vibrations for different denominations; one vibration for Rs. 5, two vibrations for Rs. 10, three for Rs. 20, four for Rs. 50, five for Rs. 100, six for Rs. 200, seven for Rs. 500 and eight for Rs. 2,000. If the app fails to recognize the denomination, it triggers a prolonged vibration and prompts the user to rescan the note. 

Acknowledging the varied needs of our users, the application includes an additional feature, “No impairment”, catering to partially or color-blind individuals. This feature allows users to opt out of the TalkBack functionality, providing a simpler and more flexible interaction method, regardless of their visual capabilities. 

The app simplifies user interaction by enabling support requests and feedback submissions via SMS or missed calls, both seamlessly embedded within the app and managed with expertise by Team Daffodil in the backend. To streamline this process, a dedicated dashboard was developed, allowing the support team to monitor and address reported issues and suggestions. These valuable inputs are utilized to continually enhance the application’s functionality.