Gas Leak Segmentation Comparison Using Different Activation Function on Fully Convolutional Network

Abstract

This paper explores the performance of a fully convolutional network, specifically the performance of differing activation functions, on a network on its ability to segment gas leakage images for gas detection. Quality management and safety control is an integral part in preserving workplace safety. This is especially true if the workplace deals with hazardous and dangerous materials, such natural-gas processing plant and chemical plant. In such working environment, one of the safety measures that can be taken is by having early detection of any possible leak. This paper tries to use the images of gas leak from thermal camera to train a semantic segmentation network to classify regions with gas leakage. Since the dataset requires videos recorded using thermal camera of gas leakage, collecting real life data has its own barriers (safety reason, availability, etc.). To help supplement the lack of available data, transfer learning on smoke video which share similar characteristics with the gas leakage video. By changing the last activation function on the fully convolutional network, we observed a difference in their performance.