基于无人机航拍与改进YOLOv3模型的云杉计数

Spruce counting method based on improved YOLOv3 model in UAV images

为解决目前苗木计数由人工完成而导致的成本高,效率低,计数精度不能得到保障的问题,该研究以自然环境下的云杉为研究对象,以无人机航拍云杉图像和拼接后完整地块云杉图像为数据源,根据云杉尺寸差异大和训练样本小的特点提出一种基于改进YOLOv3模型的云杉计数模型。该模型将密集连接模块和过渡模块引入特征提取过程,形成Darknet-61-Dense特征提取网络。通过694幅无人机航拍云杉图像测试表明,密集连接模块和过渡模块可解决YOLOv3模型小样本训练过拟合问题和云杉特征丢失问题,改进YOLOv3模型可以快速准确实现云杉计数,在精确率P、召回率R、平均精度AP、平均计数准确率MCA和平均检测时间ADT这5个评价指标上达到96.81%、93.53%、94.26%、98.49%和0.351 s;对比原有YOLOv3模型、SSD模型和Faster R-CNN模型,精确率P分别高2.44、4.13和0.84个百分点。对于拼接后完整地块云杉图像,改进YOLOv3模型的5个评价指标的结果分别为91.48%、89.46%、89.27%、93.38%和1.847 s;对比原有YOLOv3模型、SSD模型和Faster R-CNN模型,精确率P分别高2.54、9.33和0.74个百分点。该研究为利用无人机快速准确统计苗木数量的关键步骤做出有益的探索。

Densely planted seedlings in the nursery with overlapping canopies and large differences in size,have made counting seedlings difficult,if performed manually.Inaccurate seedling counting usually causes a nursery manager to make decisions mismatching with the existing state,thereby resulting the losses.It is necessary to develop automatic techniques for seedling counting,further to avoid the loss that caused by inaccurate seedling counting.In this study,the spruce plant images were collected by the unmanned aerial vehicles(DJI Phantom 4),while,a spruce image dataset was constructed.558 spruce plant images with diversity were selected,and 20 complete plot images were stitched using Pix4D mapper software.In images,the contrast,angle,and size were adjusted to expand spruce images to 4 times of original images.The training set and test set were built according to the ratio of 7:3.An improved YOLOv3 model can quickly and accurately detect the targets with large size differences,such as spruce.However,in a small sample of spruce plants,the training process was prone to overfitting,where only a few dimensional features were used in the feature extraction process,resulting in the loss of spruce feature information and less counting accuracy.The YOLOv3 model was also verified.1)Densely connected module was added to feature extraction network of YOLOv3 model,and the transfer and reuse of spruce features were strengthened.The number of model parameters was reduced to suppress the overfitting problems;2)Transition module was added to feature extraction network of improved YOLOv3 model.The spruce feature information was extracted and fused using filters with different sizes and pooling operations to avoid spruce feature loss.Five evaluation indicators including precision,recall,average precision,mean counting accuracy,and average detection time were used to evaluate the counting.Five evaluation indicators in the improved YOLOv3 model were 96.81%,93.53%,94.26%,98.49%,and 0.351 s,respectively.The improved YOLOv3 model can quickly and accurately realize spruce counting.Compared with the original YOLOv3 model,SSD model,and Faster R-CNN model,the improved YOLOv3 has significant advantages in 5 evaluation indicators.In the spruce images of a complete plot after stitching,five evaluation indicators in the improved YOLOv3 model were 91.48%、89.46%、89.27%、93.38%,and 1.847 s,respectively.Compared with the original YOLOv3 model,SSD model,and Faster R-CNN model,the performance of new model has significantly improved.The results demonstrated that the counting result of improved YOLOv3 model was greatly optimized,and further to make a useful exploration for UAV.