Visual Early Leakage Detection for Industrial Surveillance Environments
Chengang Lyu, Yage Liu, Xuekai Wang, Yuxin Chen, Jie Jin, Jiachen Yang
Abstract
Liquid leakage can cause industrial accidents. Current liquid leakage detection methods judge the leakage state by analyzing signals from special intrusion sensors, which is not ideal for early leakage because of sensor-limited sensitivity. Visual information from surveillance systems deployed in industrial environments can reflect early leakage that cannot be monitored by such pressure sensors. In this article, we propose a visual early leakage detection system based on a visual background extractor (Vibe) and EfficientNetB0. First, we extract the translucent and small potential leakage candidates based on Vibe, which include leakage targets and environmental interference. Then, to further recognize leakage targets in potential leakage candidates, we explore convolution neural network (CNN) models and a few recently proposed methods, and compare them with two different evaluation criteria. Our model based on EfficientNetB0 performs best and achieves 99.526% accuracy. In addition, our CNN model for leakage recognition with a smaller size is feasible for industrial applications. Experiments are conducted on the leakage dataset from the surveillance video from the Tianjin Binhai Heating Station and the detection results are consistent with real leakage situations. Our leakage detection system has high sensitivity and accuracy, which meets the requirements of early leakage detection.