YOLO-FSD: A Deployment-Validation-Oriented Lightweight Fire Smoke Detection Network for Resource-Constrained ZYNQ7020 FPGA Edge Platforms

Abstract

Early-stage flame and smoke in forest fire scenes are often small, weakly textured, and easily confused with complex backgrounds, while many accurate YOLO-based detectors are difficult to deploy on resource-constrained FPGA devices. This study proposes YOLO-FSD, a lightweight detector derived from YOLOv4-tiny for FPGA-oriented deployment. It integrates inverted residual and depthwise separable structures, a lightweight semantic enhancement (LSE) block at the deep feat2 feature, a lightweight P4 detection head, and a shallow detail compensation branch. On the combined test set of the D-Fire and New Fire and Smoke datasets, YOLO-FSD achieves a mean average precision at an intersection-over-union threshold of 0.5 (mAP50) of 69.36%, with 3.951 M parameters and 1.520 G multiply-accumulate operations (MACs). Compared with YOLOv4-tiny, mAP50 increases by 2.76 percentage points, while parameters and MACs decrease by 32.76% and 55.53%, respectively. For deployment validation, batch normalization (BN) fusion and 16-bit integer (INT16) parameter conversion are followed by fixed-point forward inference on a Xilinx Zynq-7020 FPGA, with a software-side parameter-quantization sensitivity analysis used as an intermediate check. FPGA raw output evaluation achieves 69.14% mAP50, only 0.22 percentage points below the PyTorch 32-bit floating-point (FP32) model, demonstrating the feasibility of FPGA-side forward inference.