DK-YOLO: a fire detection framework with noise suppression and dynamic feature extraction

Fire disasters cause significant economic losses and casualties globally, with the increasing complexity of environments intensifying the need for efficient fire detection systems. This study presents the DK-YOLO model, which addresses challenges in fire smoke detection through noise reduction and dynamic feature extraction. The model incorporates an attention-guided denoising convolutional network (ADNet) to suppress noise and improve image quality, and replaces traditional convolutions with adaptive kernel convolution (AKConv) for enhanced feature representation of dynamic flames and smoke. Experimental results show substantial improvements in mean average precision (mAP), particularly in noisy and dynamic scenarios. This work highlights the effectiveness of ADNet and AKConv in fire detection and provides a scalable framework for real-time disaster warning systems.