复杂大田场景中麦穗检测级联网络优化方法  

Optimization Method for Cascade Network of Wheat Ear Detection in Complex Filed Scene

作　　者：谢元澄[1] 何超 于增源 沈毅[1] 姜海燕[1,2] 梁敬东[1] XIE Yuancheng;HE Chao;YU Zengyuan;SHEN Yi;JIANG Haiyan;LIANG Jingdong(College of Information Sciences and Technology,Nanjing Agricultural University,Nanjing 210095,China;National Engineering and Technology Center for Information Agriculture,Nanjing Agricultural University,Nanjing 210095,China)

机构地区：[1]南京农业大学信息科学技术学院,南京210095 [2]南京农业大学国家信息农业工程技术中心,南京210095

出　　处：《农业机械学报》2020年第12期212-219,共8页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2016YFD0300607);江苏省重点研发计划(现代农业)重点项目(BE2019383);中央高校基本科研业务费自主创新重点项目(KYZ201550、KYZ201548)。

摘　　要：单位种植面积的麦穗数量是评估小麦产量的关键农艺指标之一。针对农田复杂场景中存在的大量麦芒、卷曲麦叶、杂草等环境噪声、小尺寸目标和光照不均等导致的麦穗检测准确度下降的问题,提出了一种基于深度学习的麦穗检测方法(FCS RCNN)。以Cascade RCNN为基本网络模型,通过引入特征金字塔网络(Feature pyramid network,FPN)融合浅层细节特征和高层丰富语义特征,通过采用在线难例挖掘(Online hard example mining,OHEM)技术增加对高损失样本的训练频次,通过IOU(Intersection over union)阈值对网络模型进行阶段性融合,最后基于圆形LBP纹理特征训练一个SVM分类器,对麦穗检出结果进行复验。大田图像测试表明,FCS RCNN模型的检测精度达92.9%,识别单幅图像平均耗时为0.357 s,平均精度为81.22%,比Cascade RCNN提高了21.76个百分点。The number of wheat ears per planting area is one of the key agronomic index to evaluate wheat yield.In the field scene,there are usually great differences in the shape,size and posture of wheat ears,and there are serious occlusion between leaves and ears and between ears and ears.At the same time,wheat awn,curly wheat leaves,weeds and uneven illumination introduced a lot of background interference.These complex factors led to a high false detection rate in traditional methods based on color and texture features.The detection method based on deep learning has a high missed detection rate for small-size rice ear images in practical application.To solve these problems,a wheat ear detection method FCS RCNN based on deep learning was proposed.Taking Cascade RCNN as the basic network model,a feature pyramid network(FPN)was introduced to fuse shallow detailed features and high-level rich semantic features,and online hard example mining(OHEM)technology was added to increase training frequency for high-loss samples,the network was fused by the IOU threshold.Finally,a SVM classifier was trained based on the circular LBP texture features to carry out the reinspection of wheat ear detection results to further reduce the detection error.In the wheat field image test,the detection accuracy of FCS RCNN model reached 92.9%,the average precision(AP)was 81.22%,the average time to identify a single image was 0.357 s,and the AP was 21.76 percentage points higher than that of the original Cascade RCNN.The results showed that the proposed method had better detection results for wheat ear detection in complex scenes,which provided a new idea for wheat yield estimation.

关 键 词：麦穗计数 目标检测 Cascade RCNN IOU级联 复杂场景 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]