Coin Sorting Machine using ML & AI

Tools & methods

• Raspberry Pi 4B
• USB webcams
• Python
• OpenCV
• TensorFlow / Keras
• TensorFlow Lite
• EfficientNet-B0
• pigpio
• SolidWorks

Project overview

We developed an automated coin sorting system for Sri Lankan LKR 1, 2, 5, and 10 coins as Group 12's ME 5212 project at the University of Ruhuna. The prototype brings together mechanical handling, computer vision, embedded inference, and servo actuation.

The aim was to replace denomination decisions based only on size and weight with image-based classification using coin appearance, surface texture, color, and embossed features. A Raspberry Pi 4B connects recognition to the physical sorter.

My contribution & the team

I trained the machine learning model, developed Raspberry Pi/Python code with Janith Rukshan, and contributed to fabrication with the full team. The report's role table distinguishes these responsibilities from the mechanical design, electronics design, and initial camera configuration.

Mechanical design & sorting sequence

The mechanism separates coin feeding, imaging, rotation, and routing. A container and vertical tube align the coins; the pusher and holder control the coin's position before release. A carrying rotator exposes the second side, and a directing slider sends the classified coin toward its collection bin.

1. Load bulk coins into the container and arrange them vertically.
2. Capture the first side of the bottom coin using camera 1.
3. Push the coin into the carrying rotator.
4. Rotate the carrier halfway and capture the other side with camera 2.
5. Rotate and release the coin into the directing slider.
6. Position the slider for the denomination's bin.

SolidWorks supported the mechanical design. The team manually cut and assembled accessible materials because precision 3D printing was outside the project budget.

Dataset & machine learning

The team collected more than 3,000 images of LKR 1, 2, 5, and 10 coins. Coins sourced from a bus driver supplied real circulation wear, dirt, scratches, and reflections. Images covered different lighting conditions; rotation, scaling, and brightness or contrast adjustments expanded the training data.

EfficientNet-B0 provided the classification architecture. The report describes TensorFlow/Keras training on Google Colab with GPU acceleration, using the Adam optimizer and categorical cross-entropy loss. The trained model was converted to TensorFlow Lite for inference on the Raspberry Pi.

Model development is recorded in the report's Liner screenshots, including dataset classes, loss curves, validation accuracy, and class-wise scores.

Image processing & inference

The report describes OpenCV preprocessing with grayscale conversion, Gaussian smoothing, Canny edges, contour detection, and color segmentation. These techniques support coin boundaries, shape, texture, and surface-feature analysis.

The included runtime snippet captures from two USB webcams at device indices 0 and 2. Its prediction function converts an OpenCV frame from BGR to RGB, resizes it to 224 × 224 pixels, creates a float32 batch, invokes the TensorFlow Lite interpreter, and selects the class using the largest output score.

Electronics & servo control

A Raspberry Pi 4B runs the recognition and control sequence. Two USB webcams provide images, and three servo-control outputs connect software decisions to the coin handling mechanism. The code uses pigpio for pulse-width control; it also lists NumPy, OpenCV, Pillow, TensorFlow Lite, and gpiozero imports.

The helper maps a commanded angle to a servo pulse width and pauses after movement. In the prototype, camera capture, mechanical positioning, and actuation timing had to work together so the identified coin reached the intended bin.

Results & prototype validation

The final report records a completed, working prototype that classifies denominations and uses servo-controlled routing to place coins in their bins. It includes physical build photographs and sample classifications for all four denominations.

The report does not specify the validation-set size or split, coins per minute, or a counted end-to-end sorting success rate. The displayed 100% model scores therefore describe that development run, rather than a measured reliability figure for the whole machine.

Engineering challenges & solutions

Budget, camera compatibility, image quality, and mechanical durability shaped the final implementation. The team adapted the build through procurement, testing, and redesign.

The final USB-camera arrangement enabled the project to proceed, with capture delay remaining a tradeoff. Worn coins, reflections, and inconsistent positioning reinforced the need to develop the visual model and physical handling together.

Future development & applications

The prototype demonstrates a foundation for small-scale coin handling in banking, vending, retail, and self-service systems. The report proposes several next steps:

• Improve acquisition with Raspberry Pi cameras and refine coin positioning with more precise fabricated or 3D-printed parts.
• Optimize inference or evaluate a more capable embedded processing platform.
• Expand the dataset for additional denominations and currencies.
• Add a GUI for denominations, coin totals, and sorting statistics.
• Explore IoT monitoring, transaction integration, and continuous feeding with a modular conveyor.

These are proposed extensions to the completed prototype; the report's demonstrated scope remains the four-denomination sorter.

System at a glance

An overview of the documented project scope.

1. 01Feed & align
2. 02Image first side
3. 03Push & rotate
4. 04Image second side
5. 05Classify denomination
6. 06Direct to bin

Images, videos & documents

Final project report · ME 5212 · Group 12 (PDF) Document ↗

Project resources

The final report is the primary record for the design, implementation, role allocation, figures, and results shown on this page.