摘要
针对航拍林区虫害图像的虫害区域不规则和传统识别方法泛化能力差的问题,提出一种基于全卷积神经网络(Fully convolution networks,FCN)的虫害区域识别方法。采用八旋翼无人机航拍虫害林区、获取林区虫害图像,并对虫害区域进行像素级标注,用于模型训练;将VGG16模型的全连接层替换为卷积层,并通过上采样实现端到端的学习;使用预训练的卷积层参数,提升模型收敛速度;采用跳跃结构融合多种特征信息,有效提升识别精度,并通过该方法构造了5种全卷积神经网络。试验表明,针对林区航拍虫害图像,FCN-2s在5种全卷积神经网络中区域识别精度最高,其像素准确率为97. 86%,平均交并比为79. 49%,单幅分割时间为4. 31 s。该方法与K-means、脉冲耦合神经网络、复合梯度分水岭算法相比,像素准确率分别高出44. 93、20. 73、6. 04个百分点,平均交并比分别高出50. 19、35. 67、18. 86个百分点,单幅分割时间分别缩短47. 54、19. 70、11. 39 s,可以实现林区航拍图像的虫害区域快速准确识别,为林业虫害监测和防治提供参考。
Aiming at the problem of irregularity of the pest area in the aerial images taken over forest area(discussed as forestry pest images in the following)and the poor generalization ability of traditional recognition method,a method for pest image segmentation based on full convolution network was proposed to realize automatic recognition of pest area.Firstly,the insect image of the forest area was needed to be obtained by using the eight-rotor UAV aerial photograph technique over the pest forest area,and the pest area was marked with pixels for model training.Secondly,the full connection layer of the VGG16 model was replaced with the convolutional layer,and an end-to-end study was used by implementing up sampling;and then the pre-training convolutional layer parameters were employed to improve the convergence speed of the model;finally,the skip layer was used to fuse a variety of feature information,which effectively improved the recognition accuracy,and five convolutional networks was constructed by this method.Experiment results showed that FCN-2 s had the highest recognition accuracy among the five full-convolution networks for forestry pest images.The pixel accuracy of the segmentation results was97.86%,the mean crossover ratio was 79.49%,and the segmentation time for single image was 4.31 s.Compared with K-means,pulse coupled neural network and composite gradient watershed algorithm,its pixel accuracy was higher by 44.93,20.73 and 6.04 percentage points,respectively,the mean intersection over union towered above 50.19,35.67 and 18.86 percentage points,and its segmentation time for single image was reduced by 47.54 s,19.70 s and 11.39 s,respectively.This method can realize the rapid and accurate recognition of pest area in aerial forest areas,which provide a basis for pest detection and prevention in forest areas.
作者
刘文定
田洪宝
谢将剑
赵恩庭
张军国
LIU Wending;TIAN Hongbao;XIE Jiangjian;ZHAO Enting;ZHANG Junguo(School of Technology,Beijing Forestry University,Beijing 100083,China;Key Laboratory of State Forestry and Grassland Administration for Forestry Equipment and Automation,Beijing 100083,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2019年第3期179-185,共7页
Transactions of the Chinese Society for Agricultural Machinery
基金
中央高校基本科研业务费专项资金项目(2017JC14
2016ZCQ08)
关键词
林业虫害监测
航拍
图像识别
全卷积神经网络
语义分割
迁移学习
forest pest monitoring
aerial photography
image recognition
fully convolutional network
semantic segmentation
transfer learning
