The Multi-Scale Depth-Separable Convolution Network for Fire and Smoke Detection

Fire is one of the major disasters threatening public safety and social development. The existing methods have made some advancements in the fire and smoke detection but still face several challenges. The flame characteristics are not obvious to be detected in early stage of the fire, partially overlapping flame targets are easy to miss detection, and some fire and smoke images are difficult to identify. Aiming at the above problems, we propose the multi-scale depth-separable convolutional net (MDCNet) for fire and smoke detection. Firstly, we propose the multi-scale depth-separable convolutional (MDC) module to learn the detailed features of fire and smoke better. Secondly, we design the soft filtering mechanism (Soft-DNMS) to more accurately identify overlapping targets. Lastly, we use the confidence loss (focal loss) to improve the detection rate of difficult targets. Experiments show that MDCNet outperforms other mainstream target detection algorithms in the fire and smoke detection, as compared to the optimal YOLOv7, the mean average precision improves by 1.7%. It unequivocally demonstrates MDCNet's prowess as a potent tool for fire and smoke detection, significantly outperforming comparable methods and thereby contributing significantly to the enhancement of public safety and social development.