Deep Learning-Based Fire Detector Robust to Smoke–Fog Ambiguity in Outdoor Scenes

In previous studies, fire detection models that only differentiate between fire and smoke are presented. However, a high false detection rate occurs because of confusion between smoke and fog. In this study, a novel method is proposed to effectively distinguish between smoke and fog. A custom dataset is introduced for detecting fire, smoke, and fog, which offers a novel labeling technique to reduce the misclassification of smoke and fog. A new architecture is proposed that is aimed at enhancing the detection performance. The latest You Only Look Once version 11 (YOLOv11) is used to establish a performance baseline. Further research then focuses on improving this benchmark. To ensure accurate texture differentiation, the dataset is designed to exclude overlapping ground-truth boxes, enabling the trained model to determine object boundaries independently, which is a design approach not found in previous studies. The model design aims to maximize performance and cost efficiency. The performance is improved by adding an image pyramid layer to the existing model to improve large-object detection. Cost efficiency is improved by designing a new module, C5Go, to mitigate the additional computational load introduced by the added pyramid level. Comparative experiments on the proposed custom fire dataset demonstrate that the proposed model improves the detection performance while keeping the additional computational overhead modest. The experimental results show that the proposed model achieves a 12.86% improvement in smoke detection performance compared with YOLOv11 and attains an overall mean average precision (mAP)@50 score of 0.906, reflecting superior performance. The contributions of this work are as follows: (i) we design a three-class dataset comprising fire, smoke, and fog so as not to cause false detections; (ii) we propose a new model structure to improve performance; and (iii) we verify that the proposed method indeed improves performance through the experimental results.