Bayesian Hyperparameter Optimization of a CBAM–BiFPN-Enhanced YOLOv11-Nano Architecture for Fire and Smoke Detection

Author

Gwamzhi Ponsah Emmanuel, Prof. Jibril Danladi Jiya, Prof. Ya’u Haruna Shuaibu, Dr. Hassan Aliyu


Abstract

 Vision-based fire and smoke detection systems must balance detection accuracy with the tight computational budgets of edge deployment. This paper presents an enhanced YOLOv11-Nano architecture in which a Convolutional Block Attention Module (CBAM) is embedded within every C2f block of the backbone and fused into a Bidirectional Feature Pyramid Network (BiFPN) neck, so that channel- and spatial-attention refinement operates at both the feature-extraction and multi-scale fusion stages. Because architectural additions of this kind introduce new sensitivity to training hyperparameters, a 19-parameter Bayesian hyperparameter optimization (HPO) search was conducted with Optuna across learning-rate, regularization, loss-weighting, and augmentation dimensions, using a resilient, checkpoint-resumable trial-management framework built around a persistent SQLite study store. On a curated corpus of 20,523 fire, smoke, and negative images (25,000 bounding boxes per class), the optimal configuration (AdamW, lr0 = 0.0006, box = 0.0658, cls = 0.809, mosaic = 0.948) achieved a validation mAP50 of 0.8118 across ten completed trials, an improvement over the weakest trial by more than 82% and over the median trial by roughly 8%. When the selected configuration was used to train the full CBAM–BiFPN architecture to convergence over 400 epochs at 416×416 resolution, the model reached a baseline mAP50 of 0.8076, precision of 0.8219, recall of 0.7348, and F1-score of 0.7759, with training and validation loss curves that tracked one another closely, indicating limited overfitting. These results establish the CBAM–BiFPN–YOLOv11n baseline and its optimized training recipe as a reusable starting point for downstream refinement stages such as iterative hard-example mining and post-training quantization.



Keywords

YOLOv11; CBAM; BiFPN; hyperparameter optimization; Optuna; Bayesian optimization; fire detection; object detection



Full Text:

Download Paper PDF


References


  • Ahmed, M., Raza, S., & Khan, A. (2023). Edge AI for disaster detection: A review. IEEE Access, 11, 45678–45695. https://doi.org/10.1109/ACCESS.2023.4567890
  • Deng, X., Shi, M., Khan, B., Choo, Y. H., Ghaffar, F., & Lim, C. P. (2025). A lightweight CNN model for UAV-based image classification. Soft Computing, 29, 2363–2378. https://doi.org/10.1007/s00500-025-10512-3
  • Feurer, M., & Hutter, F. (2019). Hyperparameter optimization. In F. Hutter, L. Kotthoff, & J. Vanschoren (Eds.), Automated Machine Learning: Methods, Systems, Challenges (pp. 3–33). Springer.
  • Hao, X., Bo, L., & Fei, Z. (2022). Light-YOLOv5: A lightweight algorithm for improved YOLOv5 in complex fire scenarios. Applied Sciences, 12(23), 12312. https://doi.org/10.3390/app122312312
  • He, Y., Sahna, A., He, X., Wu, R., & Zhang, R. (2024). FireNet: A lightweight and efficient multi-scenario fire object detector. Remote Sensing, 16(21), 4112. https://doi.org/10.3390/rs16214112
  • Khan, S., Muhammad, K., Mumtaz, S., Baik, S. W., & de Albuquerque, V. H. C. (2022). Energy-efficient deep learning for fire detection in smart cities. Sustainable Cities and Society, 76, 103429. https://doi.org/10.1016/j.scs.2021.103429
  • Lan, J., Tao, Y., Wang, Z., Yu, H., & Cui, W. (2025). Light-YOLOv8-Flame: A lightweight high-performance flame detection algorithm. arXiv preprint arXiv:2504.08389.
  • Li, Y., Zhang, T., & Chen, X. (2022). Lightweight CNN for fire detection on embedded systems. Sensors, 22(3), 1056. https://doi.org/10.3390/s22031056
  • Alkhammash, E. H. (2025). A comparative analysis of YOLOv9, YOLOv10, YOLOv11 for smoke and fire detection. Fire, 8(1), 26. https://doi.org/10.3390/fire8010026
  • He, Y., Sahma, A., He, X., Wu, R., & Zhang, R. (2024). FireNet: A lightweight and efficient multi-scenario fire object detector. Remote Sensing, 16(21), 4112. https://doi.org/10.3390/rs16214112
  • Pan, W., Wang, X., & Huan, W. (2024). EFA-YOLO: An efficient feature attention model for fire and flame detection. arXiv preprint arXiv:2409.12635.
  • Wang, A., Chen, H., Liu, L., Chen, K., Lin, Z., Han, J., & Ding, G. (2024). YOLOv10: Real-time end-to-end object detection. arXiv preprint arXiv:2405.14458.
  • Xu, M., Qian, K., Zhang, H., & Hu, Y. (2021). Towards real-time fire detection on edge devices with deep learning. Fire Technology, 57(5), 2569–2590. https://doi.org/10.1007/s10694-020-01062-3
  • Zhang, Q., Xu, J., Xu, L., & Guo, H. (2023). Deep learning models for fire and smoke detection: A comparative study and a novel dataset. Fire Safety Journal, 138, 103799. https://doi.org/10.1016/j.firesaf.2023.103799

Share your valuable work from Social Media Buttons