基于改进YOLOv8的轻量化火灾检测算法

Lightweight Fire Detection Algorithm Based on an Improved YOLOv8

摘要: 本文提出了一种基于改进YOLOv8的轻量化火灾检测算法，旨在优化传统目标检测模型在火灾场景中的实时性和准确性，以满足复杂环境下的高效监测需求。针对YOLOv8算法在资源受限设备上的部署挑战，主要进行了以下两方面的创新工作：(1) 设计了FasterC2f模块，作为YOLOv8算法中C2f模块的有效替代。该模块通过创新的降维策略，对输入特征图进行高效处理，显著降低了模型的参数总量与计算复杂度。(2) 提出了一种轻量化的检测头架构——Lightweight Shared Convolutional Detection (LSCD)，进一步提升模型的检测效率和精度。LSCD通过引入共享卷积机制，有效减少了检测头的参数量，同时增强了特征图之间的全局信息融合能力，确保了即使在复杂多变的火灾场景中也能维持较高的检测精度。实验结果表明，本文提出的算法在火灾检测任务上取得了优异的表现，不仅在模型体积和计算量上实现了显著轻量化，而且保持了与原版YOLOv8相当的甚至更高的检测精度。

Abstract: This paper presents a lightweight fire detection algorithm based on improved YOLOv8, aiming to optimize the real-time and accuracy of traditional target detection models in fire scenarios to meet the requirements of efficient monitoring performance in complex environments. In view of the deployment challenges of the YOLOv8 algorithm on resource-constrained devices, the following two innovative works are mainly carried out: (1) the FasterC2f module is designed as an effective alternative to the C2f module in the YOLOv8 algorithm. Through the innovative dimension reduction strategy, this module efficiently processes the input feature graph, which significantly reduces the total amount of parameters and the computational complexity of the model. (2) A lightweight detection head architecture called Lightweight Shared Convolutional Detection (LSCD) is proposed to further improve the detection efficiency and accuracy of the model. By introducing the shared convolution mechanism, LSCD effectively reduces the number of detection heads and enhances the global information fusion ability between feature graphs, ensuring that high detection accuracy can be maintained even in complex and variable fire scenarios. The experimental results show that the proposed algorithm achieves excellent performance in the fire detection task, not only achieving significant lightweight in model volume and computational quantity, but also maintaining even higher detection accuracy comparable to the original YOLOv8.