Chilean Coin Detection and Classification using CV-based Strategies

Real-time coin detector and counter using classic Computer Vision techniques

Coin identification and recognition systems can significantly improve the automation and efficiency of systems such as vending machines, public telephones, and coin counting machines. However, coin recognition presents a challenge in the fields of computer vision and machine learning due to varying rotations, scales, lighting conditions, and distinct surface patterns.

This project focuses on designing an efficient computer vision algorithm that is robust and invariant to rotation, translation, and scale—tailored specifically for the recognition of Chilean peso coins.

Project Objectives

This project was developed for the Computer Vision course in the Ph.D. program Applied Informatics for Health and the Environment (UTEM). Its goal is to detect, classify, and count Chilean coins in real time using only classic OpenCV tools—no deep-learning models.

The proposed solution must be able to:

• Segment image regions containing coins.
• Identify and determine the denomination of each visible coin (as long as its features are not severely occluded).

Matching of Chilean peso coins.

Features

• 📸 Real-time processing from a webcam (720 p).
• 🔵 Automatic metric calibration via a standard credit card (85.6 mm × 53.98 mm).
• 🟢 Hough Circle Transform for initial detection.
• 🟡 Heuristic filtering by circularity, diameter (mm) and HSV hue for each denomination.
• 🔴 False-positive rejection (buttons, tokens) by searching for inner circles.
• 📝 On-screen overlay with total and per-denomination counts ($1 / $5, $10, $50, $100, old $100, $500).
• 🗂️ Test set covering various backgrounds (bkx, wx, blurx), a fake coin (pkx), and many coins randomly situated (mcx).

Coin Detection Results

ratio = predicted/real

Table 1 – Overall Counting Summary

Denomination	Predicted	Ground-truth	Error	Ratio
$1 / $5	34	35	−1	0.97
$10	96	183	−87	0.52
$50	60	139	−79	0.43
$100	82	174	−92	0.47
$100 (Old)	33	0	+33	—
$500	106	154	−48	0.69
?? (Not Coin)	237	23	+214	10.30

Table 2 – Per-Denomination Performance Statistics (51 images)

Denomination	Mean Ratio	Std. Dev.	Median	Accuracy
$1 / $5	0.91	0.77	1.00	51.1 %
$10	0.68	1.38	0.50	13.7 %
$50	0.43	0.38	0.33	7.8 %
$100	0.43	0.32	0.50	7.8 %
$100 (Old)	1.00	0.00	1.00	100.0 %
$500	1.21	1.30	1.00	19.6 %

Project Organization

.
├── dataset                                     : Contains all coins images
├── video                                       : Contains a video example
├── CMakeLists.txt                              : CMake instructions list
├── cam_calibration.py                          : Calibration for webcam script
├── opencv-coin-detection-and-counting.ipynb    : Test notebook
└── README.md                                   : Project Report

Requirements

Package	Min version
Python	3.9
OpenCV	4.5
NumPy	1.19
Matplotlib (optional, debugging)	3.4

Tip: Use a virtual environment (venv or conda).

Installation

git clone https://github.com/fespinozav/Chilean-Coin-Detector.git
python -m venv .venv
source .venv/bin/activate

Launch App

python chilean_coin_detection.py

# Press 'q' to close/exit