Analisis Eksperimental Penggantian Backbone YOLOv5 dengan EfficientNet-B4 pada Sistem Deteksi Nilai Mata Uang

Abstract

YOLOv5 is a variant of the You Only Look Once (YOLO) algorithm, which is widely recognized for its ability to perform fast and efficient object detection in a single-stage processing framework. The performance of YOLOv5 is strongly influenced by its backbone architecture, which is responsible for extracting visual features from input images. Consequently, backbone replacement is commonly employed as an experimental approach to analyze the impact of feature extraction variations on object detection performance. This study aims to conduct an experimental evaluation of replacing the YOLOv5 backbone with the EfficientNet architecture, specifically EfficientNet-B4, for a rupiah currency value detection system. The experiment was carried out by comparing the baseline YOLOv5 model with a modified YOLOv5 model incorporating the EfficientNet-B4 backbone, using a rupiah banknote dataset consisting of seven nominal classes. Model performance was evaluated using precision, recall, mAP@0.5, mAP@0.5:0.95, and inference time as evaluation metrics. The experimental results indicate that the use of EfficientNet-B4 leads to a decrease of 1.075% in mAP@0.5, while simultaneously increasing mAP@0.5:0.95 by 2.247%. This improvement suggests enhanced bounding box localization accuracy under more stringent Intersection over Union (IoU) evaluation criteria. However, the inference time increased significantly, from approximately 7 ms to 37.8 ms per image. Overall, these findings indicate that backbone replacement provides different performance characteristics and needs to be tailored to the specific requirements of the target application.