YOLO-MP: A lightweight forest fire detection model

Traditional methods for forest fire detection face challenges, including low efficiency, high cost, and significant susceptibility to environmental factors. Furthermore, existing deep learning approaches exhibit deficiencies in feature extraction and model lightweighting capabilities. To address these challenges, this paper proposes a lightweight forest fire detection model, designated YOLO-MP. This model first introduces a novel lightweight backbone network, Ghost-HGNetV2 (GHGNet), which effectively reduces the parameter count and computational cost, suppresses noise, and concurrently enhances the model’s feature extraction capabilities. Subsequently, the Feature Pyramid Shared Convolution (FPSConv) module is incorporated for multiscale feature extraction, enabling the efficient capture of characteristic information related to forest fire targets. Following this, the Gradient Enhanced Reparameterized Block (GERB) is introduced to improve the efficiency of the lightweight model and enhance its gradient propagation capabilities. Additionally, a new loss function, Wise-Efficient IoU (W-EIoU), is designed to improve the model’s learning capacity and generalization performance across samples of varying quality. Finally, the experimental results demonstrate that YOLO-MP achieves improvements over the baseline model in recall, mAP50, and mAP50-95 by 2.76%, 1.52%, and 1.18%, respectively. The model comprises only 2.07M parameters, representing a 31% reduction compared to the baseline, and requires 6.02 GFLOPs, a decrease of 26% from the baseline. • Use the GHGNet backbone to enhance feature extraction with fewer computations. • Use FPSConv for multi-scale features and an expanded receptive field • Introduce the GERB module in the neck to enhance gradient flow and feature fusion. • Improve target localization accuracy through the use of W-EIoU loss.