Lightweight and Interpretable Coin Recognition and Counting UsingGeometric Detection and Fuzzy Score-Based Classification

Abstract

Deep learning-based coin recognition approaches typically require large, annotated datasets and substantial computational resources, yet offer limited interpretability. Such characteristics limit their applicability in lightweight, resource-constrained vision systems. Therefore, this study aims to develop and systematically evaluate a lightweight, interpretable coin recognition and counting method based on geometric detection and fuzzy-score-based classification. The main contribution of this work lies in integrating the Hough Circle Transform, contour-based circularity validation, and a weighted fuzzy score mechanism that aggregates diameter, circularity, and HSV color features without relying on data-driven model training. The proposed approach prioritizes computational efficiency and decision transparency, while maintaining robustness under varying lighting and object configurations. An experimental evaluation was performed on 40 test images containing 362 coins under both bright and dim lighting conditions, with aligned, scattered, and overlapping arrangements. The system achieved a detection rate of 87% and an object-level classification accuracy of 79%. Although image-level accuracy reached 50% under strict evaluation criteria, detailed error analysis indicates that performance degradation is primarily associated with segmentation limitations in overlapping configurations rather than instability in the fuzzy scoring mechanism. These findings demonstrate that a calibrated geometric and fuzzy-based approach can provide a transparent and computationally efficient alternative for small-scale vision applications without requiring large training datasets.