VOC gas leakage detection using infrared image and convolutional neural networks

Abstract

Leakage of volatile organic compounds (VOC) gas is one of the main sources of air pollution, and it poses a great threat to health and safety in many ways. Optical gas imaging (OGI) technique utilizes mid-wave infrared camera to visualize VOC gas and helps people observe the leakage of VOC gas. In this paper, we propose a novel method that utilizes deep learning technique and convolutional neural networks to detect the leakage of VOC gas from single-frame mid-wave infrared image. The proposed method consists of three components: color transformation pre-processing unit, feature extraction networks, and single-stage object detection sub-networks. Location-aware deformable convolution, which adjusts its sampling grid to fit the ever-changing shape of VOC gases, is employed for better feature extraction. Besides, a new loss function called leakage center loss is introduced to estimate where leakage comes from, and it forces the network to pay more attention to leakage center where the density of VOC gases is higher than dissipated parts. The proposed method is evaluated using a self-collected dataset where thousands of gas images are captured and annotated. Experimental results show that location-aware deformable convolution contributes to around 7% mAP improvement, while leakage center loss contributes to around 4% mAP improvement. Finally, our method achieves 81% mAP, which is better than existing general-purpose object detection methods. By simplifying the network architecture, our proposed method can also be implemented on embedded system for handheld VOC leakage detection devices.