由于火灾具有快速蔓延的特性和较高的破坏力，实现火灾的早期探测是十分必要的，针对火灾检测算法的研究也尤为重要。该文提出了一种改进的YOLOv8算法，通过集成轻量型模块SlimNeck和切片辅助推理方法SAHI，分别优化了YOLOv8算法的网络结构和推理框架，将火灾数据集目标分类为火焰(fire)、烟雾(smoke)和干扰项(default)。实验结果表明，SlimNeck-YOLOv8算法比相关的先进算法具有更优的火灾检测性能，与YOLOv8模型相比，查全率(recall)增长了2.7%、平均精度(mAP)增长了0.2%，检测速度提高了35 fps，同时也降低了计算负担。在SlimNeck-YOLOv8基础上进一步优化推理框架所得的SlimNeck-YOLOv8+SAHI算法，有效改善了漏检与误检现象。该研究有助于提升火灾检测系统的速度和精度，为火灾预警工作提供了有力的技术支持。

Objective: With the rapid and continuous advancement of urbanization at an astonishing pace, fire accidents are happening with increasing frequency globally. A sudden fire outbreak holds a significantly high probability of causing extensive and severe harm to society. Research conducted on image-based fire detection algorithms is highly beneficial and valuable in terms of extracting the detailed morphological features of fires or smoke, aiding in effectively improving the efficiency of fire warnings. Methods: This study presents and introduces an improved version of the YOLOv8 algorithm. Initially, the neck network of the algorithm is strengthened by integrating the SlimNeck lightweight module. Then, the inference framework of the YOLOv8 algorithm is substituted with slicing-aided hyper inference (SAHI) to further enhance the capability of the algorithm to detect small targets. Moreover, fire and smoke are two crucial target categories in fire scenarios. Given the inherent complexity of fire image backgrounds, which frequently contain numerous interferences from nonfire categories, fire dataset targets are classified as fire, smoke, and default. Results: Experimental results clearly indicate that the SlimNeck-YOLOv8 algorithm showcases superior fire detection performance compared with other related advanced algorithms. In contrast to the YOLOv8 algorithm, the recall rate of this algorithm is elevated by 2.7%, mean average precision (mAP) is increased by 0.2%, and detection speed is accelerated by 35 frames/s. Simultaneously, with the developed algorithm, the computational burden is effectively reduced. Conclusions: By integrating SlimNeck and SAHI, respectively, to optimize the network structure and inference framework of the YOLOv8 algorithm, the improved YOLOv8 algorithm is utilized for detecting fire and smoke, which has, to a certain extent, remedied the shortcomings of the YOLOv8 algorithm for this purpose. To effectively verify the performance and effectiveness of the proposed algorithm, the model is not merely trained on the fire dataset but is trained on coco128 dataset under precisely the same training epochs and parameters. This is done with the specific aim of conducting comprehensive tests to accurately evaluate model performance. The improved algorithm proposed in this study has successfully achieved the expected goals of significantly enhancing the mAP, recall, and speed of the YOLOv8 algorithm for detecting fire and smoke and concurrently reducing the rates of missed and false detections. This advancement holds great promise for enhancing the reliability and effectiveness of fire detection systems, providing prior and more accurate warnings to minimize potential losses and damages caused by fires. The combination of innovative techniques and targeted optimizations presented in this research offers valuable insights and practical solutions in the fire safety field and related applications.