Enhancing Fire and Smoke Detection with YOLOv8: A Comparative Study of Self-Supervised Learning and Attention Mechanisms

In this study, we conduct a comparative performance evaluation of object detection systems using YOLOv8 for fire and smoke detection. The assessments focused on three baseline architectures: a YOLOv8s model, a SimCLR self-supervised pretrained model, and an enhanced version that adds attention modules and EfficientNet as backbones. As well, the larger YOLO11x model was included to analyze whether additional model size would lead to improvement in performance. All models underwent training and testing using their fire and smoke custom-labeled dataset, with training results evaluated based on COCO metrics of mAP50, mAP50-95, precision, recall, and F1-score. In addition to accuracy for object detection evaluations, other dimensions were compared such as training time of the object detection models used, size of the neural networks developed detected fires or smoke as well as ROI (return on investment). The experimental results demonstrated that although the SimCLR self-supervised pre-trained model has better counts on detections, it tends to perform low in average confidence score and overall F1 measure which is improved significantly relying only on attention-enhanced architecture. The YOLO11x model outperformed others in raw accuracy but incurred significantly higher computational costs. Our study illustrates incorporating attention layers along with self-supervised pretraining improves performance while maintaining a small footprint during training time.