顾及样本均衡性的松材线虫病疫木识别研究  

Study on Thinking the Negative and Positive Samples Balance for the Pine Wild Disease Tree Detect Use Deep Learning

作　　者：卢涛 LU Tao(Natural Resources Bureau of Liuzhi Special District,Liupanshui Guizhou 553400)

机构地区：[1]六枝特区自然资源局,贵州六盘水553400

出　　处：《山东林业科技》2024年第6期82-87,共6页Journal of Shandong Forestry Science and Technology

摘　　要：松材线虫病是一种快速传播的毁灭性森林病害,其防控关键在于及时开展疫木监测。针对在喀斯特山区林地中目标尺度小、干扰因素多导致候选框生成困难和样本不均衡的问题,提出一种基于IoU采样平衡的疫木检测算法SB-HEM(sampling-balance-based hard example mining,SB-HEM),并从检测精度、检测速度两个方面分别对比分析评价SB-HEM各部分的检测性能。实验表明:SB-HEM疫木检测的F1值和mAP分别为78.96%、81.66%,SB-HEM有效提高了网络模型识别负样本的能力,从而实现了松材线虫病疫木精准识别,为松材线虫病防控提供了技术参考。Pine wilt disease(PWD)is a rapidly spreading and devastating forest disease that its control lies in timely detection of infected wood.To address the problem of generating candidate boxes with small target scale and multiple interference factors in karst mountain forests,which leads to difficulty in candidate box generation and imbalanced samples,a disease detection algorithm SB-HEM(sampling-balance-based hard example mining,SB-HEM)based on IoU sampling balance is proposed.The detection performance of each part of SB-HEM is evaluated from the perspectives of detection accuracy and detection speed.The experimental results show that the F1 value and mAP of SB-HEM infected wood detection are 78.96%,81.66%,respectively.SB-HEM effectively improves the network model's ability to identify negative samples,thereby achieving precise identification of infected wood for PWD,providing technical reference for PWD control.

关 键 词：松材线虫病 Faster R-CNN 正负样本 采样平衡 

分 类 号：S763.18[农业科学—森林保护学]