Ivan Turasov

OCR on Arduino UNO Q: Real-time text reading with Edge Impulse

Webcam, Logitech® HD Pro

Edge Impulse Studio

USB Hub, Bus Powered

Hugging Face

Arduino® UNO™ Q 2GB

<html><body>There is something magical about a camera "reading" handwritten notes, signs, or product labels in the real world. Inspired by Edge Impulse's Two stage OCR Python example, I couldn't resist porting it to the Arduino UNO Q and Arduino App Lab.How does it work in an Arduino UNO Q? It's using Model cascading: with a lightweight detector (YOLO-Pro style or pretrained PaddleOCR) scans the frame and only wakes up the recognizer model when it finds promising text regions. This opens the door to real OCR (Optical Character Recognition) applications on edge hardware.<figure><img src="https://projects.arduinocontent.cc/8fd8432f-9e6f-4239-b6eb-2b681c92221b.png"/><figcaption>Testing the OCR Edge Impulse model cascading with the Arduino UNO Q from the Arduino App Lab</figcaption></figure>Results? It's usable (a bit slow on the UNO Q but way better than a single monolithic OCR model). This isn't just a demo — it's a blueprint for Model Cascading Edge AI pipelines that you can remix for anything that you need, such as reading license plate readers, assistive tech, inventory scanners, or even augmented reality overlays.<h3>Run your own OCR project</h3>Clone this project to your computer <a href="https://github.com/mpous/ocr-linux-arduino-applab">https://github.com/mpous/ocr-linux-arduino-applab</a>Copy this project into the Arduino UNO Q folder `/home/arduino/ArduinoApps`. I usually use the VS Studio Code Remote Host plugin to SSH into the Arduino UNO Q and move the files.Open the Arduino App Lab, and the application may appear in your Arduino UNO Q `My Apps` list. <figure><img src="https://projects.arduinocontent.cc/c2860116-54ff-4b3f-821c-c567e29c6f0a.png"/><figcaption>Arduino App Lab OCR Application</figcaption></figure>Connect your USB webcam to the powered USB hub and let's try the application.<figure><img src="https://projects.arduinocontent.cc/8bbff369-6e50-4b32-958d-a56aa42c45c2.png"/><figcaption>Two-stage OCR for Linux (Edge Impulse) README file in the Arduino App Lab</figcaption></figure>Now it's time to click the `Run` button in the top-right corner from the Arduino App Lab and see how the application gets deployed into the Arduino UNO Q.<figure><img src="https://projects.arduinocontent.cc/e9cac2d7-2c42-40f0-be97-d543249c3045.png"/><figcaption>Starting the Edge Impulse OCR Demo Application in your Arduino UNO Q</figcaption></figure>And now automatically a browser interface may open with your Arduino UNO Q local IP address with the port 5001 to visualize the application.<figure><img src="https://projects.arduinocontent.cc/ce395221-037b-4b01-a43b-1f55a3a8d02d.png"/><figcaption>The Edge Impulse OCR Demo Application</figcaption></figure>Select your camera port. In my case the camera is at `/dev/video2` and click the Start button to start testing the OCR demo application.<figure><img src="https://projects.arduinocontent.cc/c2428a04-6e4c-4aa4-baff-76221bb15748.png"/><figcaption>Testing the Edge Impulse OCR Demo Application</figcaption></figure>Inference is slower (about 1 FPS) than on more powerful SBCs, but model cascading makes it feasible where a single model would be very slow. Optimizations are welcome! Can you get faster inference? Feel free to comment and share :)</body></html>