A novel self-attention weight transformer for air pollution smoke detection

Abstract

There is a necessary combustion facility called the flare stack to ensure production safety in every petrochemical plant, smelter, and refinery all over the world. Because of the incomplete combustion of flare gas, the flare stack discharges a large amount of smoke into the atmosphere and make it dirty. The air pollution is becoming more and more serious caused worldwide concern. Hence, there is in desperate need of an efficient and available smoke detection method to protect the atmosphere. To this end, we present a novel Self-Attention Weight Transformer (SAWT) that can detect accurately smoke, then guarantee full combustion of the flare stack. First, we are concerned about the weight relationship between channels and draw on the self-attention mechanism in the MobileViT block structure to adaptively adjust channel-wise feature weight. Second, we leverage short connections to thoroughly learn and fuse local and global features. Results of experiments on a real smoke dataset reveal that the proposed SAWT achieves superior performance to the popular deep CNNs and state-of-the-art smoke detection algorithms.