CHG-YOLO: A Lightweight Forest Fire Detection Framework

Abstract Abstract: To address the challenges of substantial scale variations of flame and smoke targets, insufficient discriminability of weak-feature targets, and the susceptibility of fine-grained information to loss in complex forest environments, as well as the difficulty of existing detection models in balancing detection accuracy and lightweight design, this study proposes a lightweight forest fire detection framework, termed CHG-YOLO. The framework collaboratively optimizes the network architecture along three dimensions, namely global-local feature perception, weak-target feature enhancement, and fine-grained information preservation. The global-local feature perception layer incorporates the proposed C3k2_RBC module, which leverages large-kernel reparameterization to expand the effective receptive field and enhance global contextual perception and local detail extraction capabilities for flames and smoke. It is further combined with the ADown module to reduce the number of model parameters and mitigate feature information loss during downsampling. The weak-target feature enhancement layer incorporates the proposed HyperACE_Fire module, which enhances the feature discriminability of weak flame and smoke targets and suppresses complex background noise interference in forest environments through attention enhancement, multi-scale spatial context modeling, and a feature-adaptive fusion mechanism. The fine-grained information preservation layer incorporates the proposed GSConvE module, which reduces the loss of fine-grained information during downsampling and enhances the representation capability of fine-grained features in flame textures and smoke edges through dual-path feature extraction and channel reorganization. Experimental results on the self-built forest fire dataset show that CHG-YOLO achieves a precision of 83.2%, a recall of 77.1%, an mAP50 of 84.2%, and an mAP50-95 of 64.8%, with 2.2 M parameters, representing a reduction of approximately 8.3% compared with the YOLOv13 baseline model, and an inference speed of 375.7 FPS. Additionally, cross-dataset testing further validates the generalization stability of the proposed framework across different scenarios. Meanwhile, it achieves a favorable balance among detection accuracy, lightweight design, and inference efficiency, providing a useful reference for forest fire monitoring applications, such as UAV inspection and edge-based fire monitoring.