YOLOv5s-Contextual-Fire: Introducing Contextual Transformer to YOLO for More Accurate Fire Detection

With the increase of buildings as well as the climate change, the number of fire accidents has gradually increased in recent years, bringing huge losses to human life and property. Efficient fire detection from image is of great significance that can be used for early fire warning. Convolutional neural networks (CNNs) can extract fire-related features from images. However, due to variations in nature environments, there are significant differences in the color and shape of flames, resulting in unsatisfactory recognition performance for conventional CNNs. To address this challenge, YOLOv5s-Contextual-Fire is proposed, which integrates contextual transformer module into the neck layer of YOLOv5, enabling the proposed model to better integrate backbone network features. The proposed model achieved a precision of 76% a 62.1% recall and 67.2% AP@0.5 on the self-collected dataset. Compared to other YOLO series models, the proposed model improved by up to 3.3% AP@0.5. YOLOv5s-Contextual-Fire has shown great potential for fire detection.