YOLO-DSDFE: Dual-Stream Depthwise Feature Enhancement for Lightweight Fire and Smoke Detection

Abstract

Fire and smoke exhibit different spatial characteristics: fire regions often contain localized high responses, contour variations, and irregular boundaries, whereas smoke usually presents weak texture, blurred boundaries, and broader spatial extent. Lightweight detectors commonly apply the same kernel-size and dilation pattern to all channels within a feature stage, providing limited receptive-field diversity. This paper proposes YOLO-DSDFE, a MobileNetV2-based YOLOv4-tiny detector that inserts a Dual-Stream Depthwise Feature Enhancement (DSDFE) module at feat2. DSDFE splits channels into a local 3 × 3 depthwise branch and a cascaded context branch, then applies channel concatenation, group-2 channel shuffle, residual fusion, and ReLU6. It introduces complementary local- context processing without complex attention or an additional detection scale. On D-Fire, DSDFE improves Fire AP50, Smoke AP50, mAP50, and recall by 1.69,1.89,1.79, and 4.31 percentage points, respectively, while add only 0.006 M parameters and 0.001 G MACs. An INT16 implementation on Zynq-7020 achieves 66.30% mAP50,0.42 points below PyTorch, indicating limited numerical degradation and fixed-point compatibility.