A wildfire smoke detection based on improved YOLOv8

A novel FPN network architecture, which designed to modify and improve the performance of the original YOLOv8 mode to overcome the challenges associated with diminished detection accuracy and sluggish wildfire smoke detection, is introduced in this paper. The architecture integrates the prediction layer, downsampled four times in its feature map size from the original image, in tandem with ODConv. These augmentations aim to improve the network’s feature fusion capability and predictive proficiency. Initially, the model replaces the traditional convolution in the intermediate layer with ODConv, resulting in a substantial performance enhancement. Acknowledging the non-rigid nature of smoke and the considerable variation in target size, especially prevalent in real-world settings with smaller targets, the addition of the prediction layer, subsampled four times from the original image’s feature map size, enhances the model’s ability to capture shallow feature information. Experimental verification underscores the efficacy of the improved YOLOv8 in smoke detection, demonstrating the precision attains 91.37% while recall improves to 87.67% and the mean average precision (mAP) of 95.18% for mAP50 and 67.43% for mAP50-95, signifying enhancements of 1.29%, 4.61%, 2.15%, and 3.86% compared to the original YOLOv8. Copyright © The Author(s) 2024.