A self-attention network for smoke detection

Straw burning strongly affects atmospheric environment quality and threatens human health. Smoke is the most significant feature of straw burning, thus smoke detection is used to detect the occurrence of straw burning. Due to the diversity of smoke and the interference of smoke-like objects, it may cause false negatives and false positives of detection results, and the issue of accurate smoke detection in the actual scene still remains a challenge for existing networks. In this work, we propose a novel self-attention network for smoke detection (SAN-SD) to solve the problem. The key ideas of the framework are as follows: 1) a convolutional neural network (CNN) that combines the attention mechanism in the Transformer is proposed to improve the feature extraction capabilities of baseline network by aggregating and processing the information in feature map, 2) an improved feature fusion approach is proposed to better integrate the detailed features and semantic features of smoke, so that the false positive rate is reduced, and 3) the K-means clustering algorithm is used to formulate the proposal box which can make the network converge fast and improve accuracy. Experimental results indicate that the proposed SAN-SD achieves high performance in terms of detection accuracy, false positive rate, and computational efficiency.