A Deep Learning Approach for Oriented Electrical Equipment Detection in Thermal Images

Abstract

Due to the high precision and non-contact characteristics, infrared thermography has been widely used in equipment inspection to ensure the safety of electric power systems. A fundamental step toward automatic inspection and diagnosis is the detection of equipment in thermal images. Therefore, this paper presents a deep learning approach to detect equipment parts in real-time. Specifically, we propose a deep convolutional neural network that predicts the coordinates, orientation angle, and class type of each equipment part. A prior concerning orientation consistency between parts is also incorporated into our model to improve the prediction results. For evaluation, we construct a large image set containing various kinds of scenarios. Experiments on the data set show that our method is robust to noise, achieving 93.7% mean average precision when the intersection over union threshold is 0.5, and running at 20 fps on GPU. We believe that our high accurate detection results can benefit the subsequent diagnosis.