YOLO-SAD for fire detection and localization in real-world images

Abstract

Fire detection aims to effectively monitor fires and provide timely warnings. Accurate identification and precise localization are critical in this task. However, there are several challenges, including feature diversity, background interference, foreground clutter, and the detection of small fires, which hinder overall detection performance. To address these challenges, we propose a fire detection method called YOLO-SAD, which integrates a Swin transformer, an attention and convolution mix (ACmix) module, and a decoupled head based on the YOLO (You Only Look Once) architecture. The Swin transformer is adept at extracting more discriminative features, and thus can mitigate issues related to feature diversity. Additionally, the ACmix module partitions weights for the features, thereby reducing background interference. Finally, the decoupled head incorporates a modified loss function designed to enhance the detection of small fires. Extensive experiments have been conducted on three public benchmark datasets, and the results demonstrate the YOLO-SAD is superior to the state-of-the-art methods in terms of both qualitative and quantitative metrics. The code for this paper is available at: https://github.com/yang123456-mao/YOLO-SAD-an-improved-YOLO-based-method-for-fire-detection-and-localization