FSFYOLO: A Lightweight Model for Forest Smoke and Fire Detection

The detection and identification of forest smoke and fire are critical for forest fire prevention efforts. However, current forest smoke and fire target detection algorithms confront obstacles such as high memory usage, computational costs, and deployment difficulty. Regarding these key issues, this paper presents FSFYOLO, a lightweight forest smoke and fire detection model based on the YOLOv8s model. To efficiently extract key features from forest smoke and fire images while reducing computational redundancy, the lightweight network EfficientViT is used as the backbone network. A lightweight detection head, Partial Convolutional Head (PCHead), is designed using the shared parameters idea to greatly minimize the amount of parameters and computations by leveraging shared convolutional layers and branched processing, thus achieving the lightweight design of the model. In the neck network, a lightweight feature extraction module, C2f-FL, is built to more fully extract local features and surrounding contextual information to widen the receptive field. Additionally, a Coordinate Attention (CA) mechanism is integrated into both the backbone and neck networks to capture cross-channel information, directional awareness, as well as position-sensitive information, improving the model's capacity to precisely pinpoint fire and smoke in forests. The experimental outcomes results on our self-constructed forest smoke and fire dataset demonstrate that FSFYOLO reduces the number of parameters and computation by 47.6% and 60.9%, respectively, compared to the original model, while improving precision, recall, and mAP50 by 1.3%, 1.0%, and 1.0%, respectively. This demonstrates that FSFYOLO strikes a good compromise between model lightweighting and detection accuracy.

@article{HUANG2024,
  title     = {FSFYOLO: A Lightweight Model for Forest Smoke and Fire Detection},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {11},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Yinglai HUANG and Jing LIU and Liusong YANG},
  doi       = {10.14569/IJACSA.2024.0151184},
  url       = {https://doi.org/10.14569/IJACSA.2024.0151184}
}