EFS-DETR: An Efficient Fire and Smoke Detection Network Based on Frequency-Aware Feature Fusion

Abstract

Early fire warning is critical for ensuring public safety and reducing property losses. However, current mainstream DETR detectors fail to (1) reliably detect fire and smoke in challenging scenarios, particularly with diffuse, low-contrast smoke boundaries, and (2) achieve efficient deployment for real-time applications due to their high computational cost and significant inference latency. In this paper, we propose an Efficient Fire and Smoke Detector based on an end-to-end structured framework for the aforementioned challenges. First, we design a lightweight component, namely, SwiftNet, for feature extraction. SwiftNet provides efficient multi-scale feature extraction through a lightweight star-shaped gated architecture. It couples with a Transformer architecture to effectively meet the real-time requirements. Second, we introduce an Adaptive Frequency-Aware Fusion (AFFusion) strategy, which enhances feature representation by modeling high-frequency and low-frequency characteristics via spatial filtering operations. Specifically, we supply two new blocks for AFFusion, namely SharpBlock and BlurBlock, where the former enhances structural details by extracting Laplacian-based high-frequency responses and adaptively strengthening edge information through a gating mechanism, while the latter captures low-frequency representations via Gaussian-guided smoothing to preserve semantic consistency. AFFusion precisely emphasizes object boundaries and effectively mitigates missing smoke edges, thereby improving detection accuracy without sacrificing speed. Experimental results on both our self-built dataset and public datasets demonstrate that EFS-DETR achieves competitive performance.