Improved Yolov8-based Approach for Fire Detection

Abstract

Fire detection holds immense importance in ensuring the safety of individuals and assets. Enhancing the accuracy of fire detection is essential. This article focuses on the utilization and improvements of Yolov8 for fire detection. The improvements encompass three aspects. Firstly, there is an enhancement to the Yolov8 backbone network by integrating the Inception module. Secondly, a comparison is made between five attention mechanisms, such as CBAM and ECA et al., with CoordAttention performing as the most effective. Lastly, the loss function is optimized through a comprehensive analysis of different types of loss functions, such as CIoU and SIoU et al., with WIoU being identified as the top-performing option. The experimental results demonstrate that the Yolov8 model with three improvements has increased mAP50 by 2.45% and mAP50-95 by 4.38% compared to the original Yolov8n. Compared to previous models, such as SSD and Yolov7 et al., the performance metrics of mAP50 and mAP50-95 have exhibited notable enhancements, hence augmenting the precision of fire detection.