Currency Recognition Using ESP32-CAM and Edge Impulse

This project is an ESP32-CAM based system that can recognize Indian currency notes like ₹10, ₹20, ₹50, and ₹500 using a TinyML model trained on Edge Impulse. The ESP32-CAM looks at the note, runs the model locally, and then lights up a specific LED based on which denomination it detects. At the same time, the result shows up on the Serial Monitor.

Everything runs on-device after deployment. No cloud calls during detection. Once the model is flashed, it works offline.

I’ll go step by step through the hardware wiring, image collection, model training, and deployment. If you’ve never used Edge Impulse with ESP32-CAM before, don’t worry. This is very doable if you follow the order.

What This Build Does

You place a currency note under a fixed ESP32-CAM.

The camera captures the image.

The trained model predicts the denomination.

A matching LED turns ON.

You also see the detected value in the Serial Monitor.

This is basically edge AI running on a ₹600 camera board.

Here’s what I used:

1. ESP32-CAM (AI Thinker version)
2. USB to Serial (FTDI) converter
3. Breadboard
4. Jumper wires
5. 4 LEDs (any colors)
6. 4 × 100Ω resistors
7. Tripod or fixed stand for the camera
8. Arduino IDE installed
9. Edge Impulse account

The ESP32-CAM doesn’t have a USB port, so the FTDI module is required for programming.

This part is important. The ESP32-CAM is picky about wiring during upload.

FTDI to ESP32-CAM

1. FTDI TX → ESP32-CAM RX
2. FTDI RX → ESP32-CAM TX
3. FTDI 5V → ESP32-CAM 5V
4. FTDI GND → ESP32-CAM GND

Make sure the FTDI jumper is set to 5V, not 3.3V.

Programming Mode

1. Connect GPIO0 to GND before uploading code
2. Remove this connection after upload is complete

If you forget this, uploads will fail.

LEDs

1. Each LED anode goes to a GPIO pin through a 100Ω resistor
2. All LED cathodes go to common GND

Each LED represents one currency denomination.

This matters more than you’d think.

I mounted the ESP32-CAM on a small tripod so the angle never changes. The note goes in the same spot every time. No tilt. No movement.

Consistency here massively improves accuracy.

The project has three big stages:

1. Image collection
2. Model training on Edge Impulse
3. Deployment and testing

Do not skip steps or jump around.

You need images for each denomination.

I recommend at least 50 images per note, but more is better. Try different lighting and slight rotations, but keep the background simple. White background works best.

Image Collection Method

Instead of downloading datasets, I collected my own images using the ESP32-CAM itself.

This ensures the model sees exactly what the camera will see in real life.

Arduino Setup

Install this library first:

1. EloquentEsp32cam

Open Arduino IDE and go to:

Examples → EloquentEsp32cam → Collect_Images_for_EdgeImpulse

Code Changes

At the top of the sketch, update:

Change this line:

to:

ESP32-CAM only works on 2.4 GHz WiFi, so make sure your hotspot supports that.

Upload the code (remember GPIO0 to GND).

After boot, open Serial Monitor at 115200 baud. You’ll see an IP address.

Paste that IP into a browser.

You’ll get a simple web page with buttons:

1. START COLLECTING
2. CLEAR
3. DOWNLOAD

Click START COLLECTING, place the currency under the camera, and let it capture images.

When done, click DOWNLOAD, name the label (like “₹50”), and save the ZIP.

Repeat this for all denominations.

Create a new project on Edge Impulse.

Upload all the image ZIPs.

Split data into training and testing.

Then go to Labelling Queue and draw bounding boxes around the notes. Keep them roughly square.

Edge Impulse can auto-suggest boxes using YOLO, which saves time.

In Create Impulse:

1. Image size: 96 × 96
2. Processing block: Image
3. Learning block: Object Detection

In the image block:

1. Enable Grayscale
2. Generate features

Check the feature explorer. You should see clear separation between classes.

Go to Object Detection.

Set training cycles and learning rate (default works fine).

Start training.

After it finishes, check the F1 score. If accuracy is bad, add more images and retrain.

In Deployment:

1. Select Arduino Library
2. Target: Espressif ESP-EYE (ESP32 240 MHz)

Build and download the ZIP.

Extract it.

Inside, find the folder ending with _inferencing.

Copy that folder into:

Documents → Arduino → libraries

Restart Arduino IDE.

Open the example sketch from the new Edge Impulse library.

In the code:

1. Comment out:

1. Uncomment:

Select the correct board and COM port.

Connect GPIO0 to GND again.

Upload the code.

After upload, disconnect GPIO0 from GND and reset the board.

Open Serial Monitor.

Place a currency note under the camera.

Within a second or two:

1. The correct LED lights up
2. The denomination appears in Serial Monitor

If detection is inconsistent, check:

1. Camera angle
2. Lighting
3. Background
4. Image quality in training set

1. Use more than 100 images per note if possible
2. Include worn or folded notes
3. Keep lighting consistent
4. Fix camera position permanently
5. Retrain with misclassified images

TinyML is very sensitive to data quality.

This ESP32 CAM currency recognition project proves you can run real computer vision on a cheap ESP32-CAM without cloud processing. Once deployed, it’s fast, private, and works offline.

You can extend this into:

1. Assistive tools for visually impaired users
2. Automated cash sorting
3. POS validation systems
4. Vending machine note detection

If you’re getting into embedded AI, this is a solid hands-on build that actually teaches you something real.