Comparative Evaluation of YOLO Models for Gauge Detection

• Franco Rivadeneira Pontificia Universidad Catolica del Peru
• Eduardo Cabrera Yi Pontificia Universidad Catolica del Peru
• Alessandro Miyahira Pontificia Universidad Catolica del Peru
• Miguel Zinanyuca Pontificia Universidad Catolica del Peru
• Francisco Cuellar Pontificia Universidad Catolica del Peru

Abstract

In various industries, accurate gauge measurement is crucial for maintaining safety and equipment integrity, particularly in hazardous environments with flammable, corrosive, or toxic substances. Ensuring precise detection is essential to avoid accidents and enhance operational efficiency. Traditional methods often struggle with technical and safety limitations in such conditions. Consequently, creating reliable gauge detection systems for hazardous environments has become a key area of innovation, demanding robust performance and compliance with strict safety standards. In this paper, we aim to train an optimal model with the best performance for detecting gauges in hazardous environments. This is achieved by comparing the latest versions of the most frequently used detection architecture, YOLO by Ultralytics. As a result, six models with different optimizers were trained per version, with the YOLOv10 model using the NAdam optimizer emerging as the best. It achieved an F1-Score of 98.2% and a latency of 4.22ms.

Keywords: YOLO, Industries, Technological innovation, Accuracy, Safety, Reliability, Standards, Robots, Accidents, Machine Learning, Gauge Detection