Abstract: To improve the accuracy of smoke and fire recognition in complex environments, this paper proposes a dual-band smoke and fire recognition method based on 3D convolution and a spatio-temporal attention mechanism. The method in this paper fuses near-infrared and visible dual-band image data and uses time-based dynamic features and spatial-based static features in the video stream to reduce missed and false alarms. Experimental results show that the algorithm in this paper achieves 99.90% smoke and fire recognition accuracy on the dual-band data set, which is better than other 3D convolution-based smoke and fire recognition algorithms, while the model has a small number of parameters and can meet the real-time inference requirements. Therefore, the use of dual-band features combined with an attention mechanism to make full use of the dynamic information in the video can effectively improve the smoke and fire recognition accuracy.

Key words: smoke and flame recognition, 3D convolution, video dynamic features, multi-spectrum