一种改进DETR的森林火灾烟雾识别模型  

An Improved DETR Model for Forest Fire Smoke Recognition

作　　者：颜谨 肖满生[1] 王瑶瑶 朱泽宇 YAN Jin;XIAO Man-sheng;WANG Yao-yao;ZHU Ze-yu(School of Computer Science and Engineering,Hunan University of Technology,Zhuzhou 412000,China)

机构地区：[1]湖南工业大学计算机学院,湖南株洲412000

出　　处：《计算机技术与发展》2025年第2期24-32,共9页Computer Technology and Development

基　　金：湖南省自然科技基金项目(2022JJ5077)。

摘　　要：针对传统的基于卷积神经网络(CNN)的森林火灾烟雾检测中,需要大量人工设计构件、对复杂森林场景中不明显的小烟雾检测能力较差等问题,提出了一种改进DETR(Detection Transformer)的森林火灾发生早期的烟雾检测模型。首先,使用DETR作为基线,将多尺度对比局部特征模块(MCCL)和密集金字塔池化模块(DPPM)集成到特征提取模块中,用于感知小烟雾或不明显烟雾的特征;然后,提出了一种边界框迭代组合方法,生成能够覆盖整个烟雾对象的精确包围盒,以提高检查精度、减少误检与漏检;最后,利用CutMix数据增强扩充森林火灾烟雾数据集,对该方法进行评价。理论分析与实验表明,改进后的模型对森林火灾烟雾的检测精度明显高于主流模型,与传统DETR模型相比,该模型的mAP(平均精度均值)提高了4.4%,AP_(50)精度提高了3.8%。Aiming at the problems of the traditional convolutional neural network(CNN)based forest fire smoke detection,such as the need for a large number of artificial design components and poor ability to detect small smoke which is not obvious in complex forest scenes,an improved DETR(Detection Transformer)smoke detection model in the early stage of forest fires is proposed.Firstly,the multi-scale contrast local feature module(MCCL)and dense pyramid pool module(DPPM)are integrated into the feature extraction module using DETR as the baseline for sensing the features of small or invisible smoke.Then,a boundary box iterative combination method is proposed to generate an accurate bounding box that can cover the entire smoke object,so as to improve inspection accuracy and reduce false detection and missing detection.Finally,CutMix data is used to enhance and expand the forest fire smoke dataset,and the proposed method is evaluated.Theoretical analysis and experiments show that the improved model has significantly higher detection accuracy of forest fire smoke than that of the mainstream model.Compared with the traditional DETR model,the mAP(mean accuracy)of the model is increased by 4.4%,and the AP_(50) accuracy is increased by 3.8%.

关 键 词：烟雾检测 DETR 多尺度特征信息 注意力机制 CutMix 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]