Smoke and Fire Detection Based on Lightweight Depthwise Separable Convolution

This paper proposes a target detection model based on depthwise separable convolutions and a lightweight attention mechanism, specifically designed for efficient smoke and fire detection on patrol robots. The model reconstructs the neck network of YOLOv5 using depthwise separable convolutions. By combining parallel depthwise separable convolutions with different dilation rates with a residual structure, it achieves a balance between multi-scale information modeling and significantly reduced computational cost. A lightweight attention mechanism module is also introduced to effectively highlight fine-grained target edge regions such as smoke wisps and flames. Experimental results show that, under similar FPS conditions, YOLOv5_DS and YOLOv5_DS_Attention, built based on the above module, improve the accuracy from 63.3% to 70.1% and the AP50 from 79.6% to 81.2%.