Enhanced fire detection in new dataset: a fine-grained and multiscale-optimized architecture

Most existing deep learning methods are confined to single-scenario detection or coarse-grained binary classification tasks (fire/nonfire), often overlooking the critical value of fine-grained fire type classification in smart city development. However, when confronting flame targets with significant scale variations, existing models face severe challenges regarding feature extraction and multiscale adaptability, impeding precise detection. Addressing these issues, we propose EFDC-YOLO, an improved framework built upon YOLOv8. Specifically, we integrate a dynamic ghost convolution module with SS2D into the backbone network and embed a strip large kernel spatial module into the neck network. By effectively capturing discriminative spatial context information and enhancing multiscale feature representation capabilities, this architecture significantly elevates the model’s recognition precision and robustness. In addition, we release a diverse dataset covering building, vehicle, and forest fires. Experimental results demonstrate that EFDC-YOLO outperforms the original YOLOv8, achieving mAP0.5 improvements of 3.2% and 1.7% on the public M4SFWD dataset and our self-constructed dataset, respectively. These results robustly demonstrate the immense potential of EFDC-YOLO for achieving high-precision, multitype fire monitoring in complex real-world scenarios. Our codes and dataset are publicly available at https://github.com/Administor123/EFDC-YOLO/