<p>To address the challenges of small target flames and target scale variation in forest fire images, a target detection method for forest fire images based on multi-scale feature extraction was studied, with YOLOv9c as the baseline model. Initially, a lightweight feature extraction module named EGI (ECA_Ghost_InceptionV2) was proposed to serve as the backbone feature extraction network, which improved the model’s feature extraction capability and operational efficiency. Second, a P2 small target detection head was introduced; meanwhile, a small target feature fusion module was added to the Neck layer, and the CARAFE upsampling operator was incorporated, enhancing the model’s ability to extract underlying feature information. Finally, to solve the problems of misalignment and scale inconsistency in the traditional IoU loss function, Inner_DIoU was introduced. This enabled the relative relationship between bounding boxes to be described more accurately and improved the precision of target detection. The improved model was validated through experiments on the DFireDataset. Results show that it achieved a detection accuracy of 79.2%, representing a 3.8% improvement compared with the baseline model, while the number of parameters was reduced by 29%, It also maintains a real-time inference speed of over 25 FPS on edge GPUs, enabling deployment in UAV-based forest monitoring systems. In addition, the model exhibits strong robustness to complex natural backgrounds and significantly reduces false alarms compared with existing methods. These findings demonstrate that the proposed model exhibits excellent performance in small target flame detection and is well-suited for the target detection task of forest fire images.</p>