Forest Fire Detection Method Based on Improved YOLOv7

The proposed method addresses the challenge of real-time and high-precision detection of early fires in complex forest environments, a limitation faced by existing object detection algorithms. The method is based on the improvement of YOLOv7. Firstly, the Coordinate Attention (CA) mechanism is introduced into the backbone network to capture the correlation of features at different spatial positions, enhancing the model's feature extraction capabilities. Additionally, CA is incorporated into the Feature Pyramid Network (FPN) to construct the C-FPN feature fusion network, establishing cross-level connections between shallow and deep networks to shorten the path between shallow and deep features and enhance the retrieval capability of deep network position information. Secondly, the Normalized Wasserstein Distance (NWD) is utilized instead of IoU in label assignment, Non-Maximum Suppression (NMS), and loss functions to improve regression accuracy. Finally, the network structure ELAN is improved to ELAN-P, introducing new partial convolutions to reduce floating-point calculations, redundant computations, and memory access, thereby enhancing the efficiency of the detection model. Experimental results demonstrate that compared to the YOLOv7 algorithm, the proposed method achieves a mean Average Precision (mAP) of 96.3%, a 4.6% improvement over YOLOv7. Additionally, the detection frames per second (FPS) reach 67.7 flames/s, enabling real-time and accurate detection of forest fires in complex environments.