EG-ReNet:A Multi-Scale Fire Detection Framework based on Frequency-Domain Residual Enhancement and Fine-Grained Perception

Forest fire visual monitoring is a safety-critical task in which early and accurate localization of smoke and flame can substantially reduce suppression costs and ecological damage. Existing methods struggle because backbone downsampling erases the high-frequency boundary signals that distinguish smoke from cloud, standard detection heads lack the directional sensitivity needed to model combustion-driven gradient anisotropy, and nascent fire targets below the minimum detection stride fall outside the perceptual range of conventional neck architectures. To address these mismatches, EG-ReNet is proposed, a lightweight detection network built around three modules each derived from the corresponding physical constraint. The Spectral Residual Enhancement Module (SREM) recovers attenuated high-frequency residuals through a multi-scale filter bank and applies a dual-domain selection mechanism to separate genuine smoke responses from spectrally similar cloud interference. The Gradient Prior Re-Parameterization Head (GPR-Head) embeds multi-directional differential convolutions with direction-adaptive weights as combustion-informed inductive biases during training and folds them into a single equivalent convolution at inference, capturing gradient anisotropy without additional computational cost. The Fine-Grained Resolution Neck (FGR-Neck) extends the detection hierarchy to a stride-4 resolution level, resolving the perceptual blind spot for nascent fire targets that fall below the standard detection stride. On the self-built FL-FFD dataset, EG-ReNet achieves 76.13% mAP@0.5 with 1.48M parameters, a 3.88% gain over the baseline at 43% fewer parameters. Validation on the public FLAME and D-Fire benchmarks yields 89.65% and 85.83% mAP@0.5 respectively, confirming consistent generalization across scenes and demonstrating that physics-conditioned design provides a viable path to accurate and lightweight wildfire detection.