Analisa Kinerja Sistem YOLOv4-Tiny Pada Pembuatan Alat Deteksi Kecurangan Jual Beli Gas LPG 3 kg

Abstract

Abstract— The 3 kg LPG distribution plays an important role in fulfilling household gas needs in Indonesia. However, challenges such as consumer fraud, where individuals may manipulate gas cylinder exchanges, can lead to inaccurate records and financial discrepancies. This research aims to develop a visual detection system capable of identifying 3 kg LPG gas cylinders to support monitoring and fraud prevention at gas distribution. The system utilizes the You Only Look Once (YOLO) object detection, with YOLOv4-Tiny selected due to its balance between detection speed and lightweight architecture. The experiments were done to evaluate the performance of Yolov4-Tiny in fraud detection by tuning the hyperparameters. The hyperparameters that were tuned are the learning rate, batch size, and dataset split ratio, which is the ratio of training, validation, and test data. The best results were obtained using a learning rate of 0.012 and a batch size of 64, yielding a detection accuracy of 81.08%, with a precision of 84%, a recall of 87%, an F1-score of 86%, and an mAP@0.50 of 91.70%. The average image processing time was recorded at 81.07ms, with a video inference speed of 18.51 frames per second (FPS), indicating the system's potential for real-time implementation, depending on hardware constraints. The findings suggest that this system can improve the accuracy of cylinder tracking. For future research, it is recommended to explore higher maximum batch settings, and additional parameters such as momentum, decay, and subdivisions can also be tested for a better understanding of their impact on model training.