Evaluasi Kinerja Pipeline YOLOv8 dan EasyOCR pada Deteksi Kode Aset Kampus

Abstract

Manual asset tracking on campus is prone to inefficiencies and human error. This study evaluates the performance of a two-stage architecture that combines the You Only Look Once version 8 (YOLOv8) model for text localization and EasyOCR for optical character recognition on physical inventory labels. The scope of the study is limited to printed asset labels (printed text) tested within the Department of Information Technology. Evaluation was conducted using stress testing methods to assess the system’s resilience to variations in lighting and perspective distortion. The results show that YOLOv8 is highly robust at localizing label areas, achieving a mean average precision (mAP@50-95) of 70.04%. During the character extraction stage, EasyOCR produced the lowest character error rate (CER) of 14.16% under ideal shooting conditions. However, extraction performance declined, with error rates reaching 30.82% under glare conditions and 48.58% at extreme tilt angles. This study concludes that the YOLOv8 pipeline is highly suitable for object localization in campus environments, but the OCR module requires more adaptive image preprocessing optimization to handle variations in environmental conditions. This research provides a practical contribution in the form of an adaptive computer vision architecture guideline and an objective benchmark for overcoming image degradation distortions in local-scale inventory management systems.