摘要
长期以来,苗圃树苗的检测与计数一直依赖于人工抽样估计,该方法效率较低且误差较大。为解决上述问题,基于图像处理和深度学习技术,针对云杉、花楸与景观榆树3种树苗提取三类标签图像,构建了树苗检测与计数的输入数据集,并利用数据增强方法将数据集扩充了15倍。基于扩充后的数据集,提出了基于YOLOv3深度神经网络的树苗检测与计数方法,并使用迁移学习的方法对网络模型进行训练与验证,从而提高了检测与计数准确率。试验结果表明:该网络模型能够有效地克服大田环境下的噪声,实现树苗的快速、准确识别,云杉树苗图像检测时间平均每幅0.681 s,花楸树苗图像检测时间平均每幅0.698 s,景观榆树树苗图像检测时间平均每幅0.697 s,MAP值达到0.825。对100幅云杉树苗图像、50幅花楸树苗图像和50幅景观榆树树苗图像进行树苗计数测试,采用人工计数结果与系统计数结果对比进行正确率评价。研究结果表明,树苗总体识别正确率达到95.2%(其中云杉树苗计数准确率为97.5%、花楸树苗计数准确率为91.9%、景观榆树树苗计数准确率为96.2%),能够满足树苗检测计数的实际要求。
For a long time,the detection and counting of most small-and medium-sized nursery trees have been dependent on manual sampling estimation,which is relatively inefficient and has a large error.With the development of target detection technology,more and more agricultural and forestry productions have begun to use deep learning technology for intelligent operation,in order to save labor costs and improve the accuracy of estimation.To solve the above-mentioned problems,based on image processing and deep learning techniques,this study extracted three types of label images for spruce saplings,sorbus saplings,and landscape elm saplings,constructed the input data set for the detection and counting of these three saplings,and expanded the data set by 15 times using the data enhancement method.Based on the expanded data set,a detection and counting method of three kinds of saplings was proposed based on YOLOv3 deep neural network,and the transfer learning method was used to train and verify the YOLOv3 network model,so as to improve the detection and counting accuracy.In this study,a small computer with its own GPU was used,so there was no need to rent or buy large computing resources for network training and detection,resulting in saving the operation cost of small-and medium-sized nurseries.The test results showed that the network model can effectively overcome the noise in the field environment and realized the rapid and accurate identification of tree saplings.The average detection time of spruce saplings,sorbus saplings and landscape elm saplings was 0.681,0.698 and 0.697 s,respectively,and the MAP value reached 0.825.One hundred spruce sapling images,50 sorbus sapling images and 50 landscape elm sapling images were tested,and the accuracy was evaluated by comparing the manual counting results with the system counting results.The results showed that the overall accuracy rate of tree seedling recognition reached 95.2%(97.5% for spruce seedlings,91.9% for sorbus seedlings and 96.2% for landscape elm seedlings),which could meet the actual requirements of tree seedling detection and counting.It can provide technical supports for the yield estimation of saplings in the forestry intelligent production.
作者
袁叙广
赵鹏
李丹
YUAN Xuguang;ZHAO Peng;LI Dan(College of Information and Computer Engineering,Northeast Forestry University,Harbin 150040,China)
出处
《林业工程学报》
CSCD
北大核心
2022年第3期174-179,共6页
Journal of Forestry Engineering
基金
国家自然科学基金(31670717)。