Recognizing Indonesian Rupiah with Computer Vision

As a Data Science graduate, I’ve always been passionate about building AI that solves real-world problems. Back during an Innovillage project I led, I developed a computer vision system designed to recognize Indonesian Rupiah.

I want to share the story behind this project, how it works, and how you can try it out yourself

The Challenge

For many of us, telling a Rp 100.000 note from a Rp 20.000 note takes a fraction of a second. But for people who are blind or visually impaired, identifying the value of currency notes can be quite difficult.

You might wonder, don’t banknotes have those textured blind codes? Yes, they do. However, the tactile blind code feature is not available on all older banknotes. Furthermore, these codes can sometimes be too soft to feel, and some people who are blind are simply not familiar with how to use them.

The primary objective of this project was to bridge that gap: to provide access to information that increases the independence of people who are blind during everyday transactions.

The Tech I Used

To solve this, I turned to Computer Vision. Specifically, this project uses the model called YOLO v11.

I used it for something called “instance segmentation” to detect and identify Indonesian Rupiah. In simple terms, instead of just telling you what is in a picture, the model acts like a smart highlighter drawing a precise, digital outline around the specific banknote it recognizes so the system knows exactly where and what it is.

To train this model, I put together a dataset of 5.000 images, split into 4.000 images for training and 1,000 for validation.

• 50% Indonesian Rupiah: A diverse collection of Rupiah banknotes, where each note is carefully labeled with its correct denomination.
• 50% Other Currencies: A wide array of banknotes from various countries, like the US Dollar, Euro, and Yen.

I used a technique called Smart Labeling. To prevent the model from incorrectly identifying foreign currency as Rupiah, the images of the non-Rupiah currencies were intentionally left unlabeled. This method teaches the model not only what Rupiah notes look like, but just as importantly, what they don’t look like. By treating foreign currencies as background noise, the model learns to completely ignore them, which significantly reduces false detections.

The Results

The smart training paid off. When put to the test, the model achieved fantastic results:

Try It Out

Building this was a great learning experience, and I want to leave the door open for anyone else who wants to learn from it or build upon it:

• Test the live demo: You can play around with the working prototype https://moneydetection.streamlit.app/
• Model: the model itself are available in my repo https://github.com/Luthfillah31/money_detection.
• Dataset: If you need the annotated images for your own computer vision project, just send me an email!