<p>Early small fires in coal mines are of critical importance for mine safety, as most incidents begin as small fire sources. These fires essentially constitute a task of small-object detection. Detecting such targets is difficult due to harsh underground conditions, including poor lighting, smoke, confined spaces, and dense obstacles. Despite the progress brought by deep learning, considerable challenges persist in small-object feature representation, spatial information reconstruction, and the efficient deployment of models on resource-constrained edge devices, preventing current methods from meeting the demands for high accuracy and real-time detection in coal mines. To address these challenges, this study proposes CMF-Net, a lightweight multi-scale feature fusion network based on YOLOv8n. CMF-Net incorporates four modules, including an enhanced small-object feature extraction module, an adaptive upsampling operator for spatial detail reconstruction, a lightweight detection head to reduce computational complexity, and an improved loss function for better localization accuracy. Experiments on a self-built Coalmine-Fire dataset and the public Fire dataset show that CMF-Net outperforms mainstream methods, achieving 30.4% in AP-small and 73.1% in mAP@50:95. With its compact design and high inference speed, CMF-Net can be efficiently deployed on edge devices, offering a promising solution for intelligent fire monitoring in underground coal mines.</p>