安全与环境学报

为了精准检测森林城镇交界区域在烟雾等复杂干扰环境下的火灾情况，提出了一种基于ViT–YOLOv8的火灾烟雾检测模型。通过收集森林城镇交界区域等典型火灾场景图像构建高质量数据集，利用图像标注工具，精确划分出明火、烟雾、其他干扰因素等三类关键目标；并将重参数化视觉转换器（Re-parameterized Vision Transformer，RepViT）、瓶颈转换器网络（Bottleneck Transformers Network，BoTNet）及幻影卷积层（Ghost Convolutional layer，GhostConv）集成到YOLOv8的主干网络，以提升复杂环境下的火灾及烟雾识别准确度和抗干扰能力。结果表明，与原检测模型相比，该模型对于各类检测目标的平均精度均值、精确率及召回率分别由96.8%、84.4%、97.5%提升至97.7%、84.9%、98.0%。该模型不仅增强了全局信息捕捉能力，还提高了局部细节的识别精度，能够更精准地识别森林城镇交界域火灾及前期烟雾迹象。

To accurately detect fire and smoke in the complex disturbed environments of the wildland-urban interface， this paper proposes an enhanced detection model based on YOLOv8， named ViT–YOLOv8. This model introduces a Re-Parameterized Vision Transformer (RepViT) module into the backbone network of YOLOv8. Unlike traditional convolutional networks that mix spatial and channel information， this module processes spatial feature extraction and channel feature fusion separately， allowing these two processes to operate independently without interference. This enhances the model’s ability to capture and utilize key features more effectively. Additionally， by leveraging structural re-parameterization， the multi-branch topology of the token mixer is merged into a single depthwise convolution layer， optimizing the allocation of computational resources and improving operational efficiency. Furthermore， a Bottleneck Transformers Network (BoTNet) module integrated with a Ghost Convolutional layer (GhostConv) is incorporated after the spatial pyramid pooling layer in the backbone network. GhostConv increases the number of channels at a low computational cost， enabling the extraction of critical information from the original feature maps. A multi-head self-attention mechanism is introduced into the last three convolutional layers， allowing the model to consistently focus on crucial contextual information within feature maps under complex environmental conditions. This significantly enhances the model’s anti-interference capability while maintaining relatively low computational complexity. In addition to collecting high-quality images of fires in the wildland-urban interface， images of strong light sources that could be mistaken for fires were also gathered as interference samples， simulating the complex lighting conditions typical of such regions and improving the model's robustness. The constructed dataset comprises 4 268 images annotated into three categories: fire， smoke， and other. After training for 650 epochs on this dataset， comparative experiments were conducted with the baseline model YOLOv8n. Experimental results demonstrate that， compared to YOLOv8n， the proposed model achieved improvements in mean average precision (M_P)， precision (P)， and recall (R)， increasing from 96.8%， 84.4%， and 97.5% to 97.7%， 84.9%， and 98.0%， respectively. Moreover， ViT–YOLOv8 not only maintained high detection accuracy for fire and smoke targets but also exhibited superior precision in identifying interference items， showcasing excellent anti-interference capability. The model improves both global information capture and local detail recognition， facilitating more accurate detection of fires and their early smoke signs in the complex environment of the wildland-urban interface.