【作者】 李博；

【导师】 张军国；

【作者基本信息】 北京林业大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 森林火灾是主要的林业灾害类型,对生态环境和人类生命财产安全造成严重威胁。烟雾是火灾发生时的显著特征,快速准确地检测出早期的小目标烟雾,对森林火灾预警和扑救工作至关重要。传统森林火灾烟雾监测方法需要将图像传输到监控中心服务器进行集中处理,存在传输延迟、数据拥堵、检测鲁棒性差等问题,难以满足监测预警的时效性。为此,本文在自建森林火灾早期烟雾数据集的基础上,提出了一种基于混合域注意力机制的多尺度早期森林火灾小目标烟雾检测模型,并将其部署在边缘端进行试验。本文主要研究工作内容如下:1.构建森林火灾早期小目标烟雾数据集。首先通过多渠道收集到3616张真实林区场景下的森林火灾烟雾图像,利用Cycle-GAN对图像背景进行季节风格转换,生成反季节图像2400张,扩充了样本数量和背景多样性。其次,采用MSRCR算法增强RGB烟雾图像的色彩,提高其细节清晰度,减少光照变化对烟雾颜色纹理信息的干扰。最后,使用基于伪标签的自监督学习算法对森林火灾烟雾图像进行自动标注,并按照VOC标准数据集格式构建了包含6016张烟雾图像的早期森林火灾烟雾数据集,为后续算法开发提供了数据基础。2.提出基于改进YOLOv5的早期森林火灾小目标烟雾检测模型M-YOLO。考虑到边缘设备有限的算力和实时性要求,本文选择参数量较少的单阶段目标检测模型YOLOv5s作为基础模型。针对小目标烟雾特点,在YOLOv5的三尺度检测头基础上,新增一个4倍下采样检测头Detect-Tiny,并调整Neck端的特征融合方式。在自建数据集上与其他主流方法进行对比实验。实验结果表明,M-YOLO模型在小目标烟雾检测方面明显优于其他实验模型,平均准确率达到了96.74%,较YOLOv5s、YOLOv8和Faster R-CNN分别提升了1.32%、0.4%、3.26%。3.提出融合混合域注意力的多尺度早期森林火灾烟雾检测模型MEDA-YOLO。首先,设计了结合高效通道注意力机制和融合空洞卷积的空间注意力机制的混合域注意力机制,使网络能够同时关注空间和通道两个维度,过滤无关信息,突出烟雾信息。其次,将混合域注意力机制嵌入M-YOLO中,得到MEDA-YOLO模型。实验结果表明,本文提出的MEDA-YOLO模型能够有效降低对云、雾等类烟目标的误报率,平均准确率达到了97.58%,较YOLOv5和M-YOLO分别提升2.16%、0.84%。4.完成本文模型在边缘设备上的部署和试验。首先,针对检测任务需求搭建了边缘端检测系统。然后,利用TensorRT推理引擎,对MEDA-YOLO模型进行加速,降低模型的参数量和计算复杂度,提高模型的推理速度。最后,将加速后的模型部署在本文所构建的边缘设备上,并按照不同量化精度进行了测试。实验结果表明,本文模型经过FP32精度的量化加速后,在平均准确率和检测速率方面达到了最佳的总体性能,分别达到95.4%和58 FPS。更多还原

【Abstract】 Forest fires are a major type of forestry disaster that pose serious threats to the ecological environment and human life and property safety.Smoke is a prominent feature of fire occurrence,and rapid and accurate detection of early small target smoke is crucial for forest fire early warning and firefighting work.Traditional forest fire smoke monitoring methods require image transmission to the monitoring center server for centralized processing,which has problems such as transmission delay,data congestion,poor detection robustness,etc.,making it difficult to meet the timeliness and effectiveness of monitoring and early warning.To this end,based on the self-built early forest fire small target smoke dataset,this paper proposes a multi-scale early forest fire small target smoke detection model based on hybrid domain attention mechanism,and deploys it on the edge device for experiments.The main research work contents are as follows:1.Construct an early forest fire small target smoke dataset.First,3616 real forest fire smoke images in real forest scenarios were collected through multiple channels,and CycleGAN was used to perform seasonal style transfer on the image background,generating 2400anti-seasonal images,expanding the sample size and background diversity.Second,the MSRCR algorithm was used to enhance the color of RGB smoke images,improve their detail clarity,and reduce the interference of illumination changes on smoke color texture information.Finally,a self-supervised learning algorithm based on pseudo-labels was used to automatically annotate forest fire smoke images,and an early forest fire smoke dataset containing 6016 smoke images was constructed according to the VOC standard dataset format,providing data basis for subsequent algorithm development.2.Propose an early forest fire small target smoke detection model M-YOLO based on improved YOLOv5.Considering the limited computing power and real-time requirements of edge devices,this paper chooses YOLOv5 s with fewer parameters as the basic model.Aiming at the characteristics of small target smoke,a 4x down-sampling detection head Detect-Tiny is added on the basis of YOLOv5’s three-scale detection head,and the feature fusion method of Neck end is adjusted.Comparative experiments with other mainstream methods were carried out on the self-built dataset.The experimental results show that MYOLO model is significantly better than other experimental models in small target smoke detection,with an average accuracy of 96.74%,which is 1.32%,0.4%,and 3.26% higher than YOLOv5 s,YOLOv8 and Faster R-CNN respectively.3.Propose a multi-scale early forest fire small target smoke detection model MEDAYOLO that integrates hybrid domain attention.First,a hybrid domain attention mechanism that combines efficient channel attention mechanism and spatial attention mechanism that fuses dilated convolution is designed,which enables the network to pay attention to both spatial and channel dimensions simultaneously,filter out irrelevant information,and highlight smoke information.Second,the hybrid domain attention mechanism is embedded into M-YOLO to obtain MEDA-YOLO model.The experimental results show that MEDAYOLO model proposed in this paper can effectively reduce the false alarm rate of cloud-like targets such as clouds and fog,with an average accuracy of 97.58%,which is 2.16% and0.84% higher than YOLOv5 and M-YOLO respectively.4.Complete the deployment and experiment of this paper’s model on edge devices.First,an edge-side detection system is built for the detection task requirements.Then,the MEDA-YOLO model was accelerated using the TensorRT inference engine to reduce the number of parameters and computational complexity of the model and improve the inference speed of the model.Finally,the accelerated model is deployed on the edge devices built in this paper and tested according to different quantization accuracies.The experimental results show that the model in this paper achieves the best overall performance in terms of average accuracy and detection rate after quantization acceleration with FP32 accuracy,reaching95.4% and 58 FPS,respectively.更多还原