基于Faster R-CNN的大田玉米雄穗识别及抽穗期判定研究

目前田间玉米雄穗数量监测主要依靠人工进行,效率低且易出错.为了实现在复杂的田间环境下对玉米雄穗自动识别和计数的任务,使用无人机平台和田间作物表型高通量获取平台采集的田间玉米顶视图像构建数据集,使用Resnet 50作为新的特征提取网络代替原始的VGG 16来优化Faster R-CNN模型.再根据表型平台所获取的高时序、连续图像,进一步使用改进后的模型对试验小区内玉米抽穗期前后20 d的雄穗数量进行监测,以此为依据进行抽穗期判定.该方法在田间作物表型高通量平台获取的图像数据测试集中类平均精度为90.14％,平均绝对误差为4.7328;在无人机平台获取的图像数据测试集中类平均精度为82.14％,平均绝对误差为9.6948.试验结果表明:该模型在田间作物表型高通量获取平台上的检测结果优于无人机平台,且具备一定的应用价值.

Research on male ear detection and tasseling stage identification of field maize based on faster R-CNN

At present, the monitoring of the number of tassel of maize in the field mainly depends on manual work, which is inefficient and error-prone. In order to realize the task of automatic identification and counting of male ear in complex field environment, this paper used the images collected by UAV platform and the Plant High-throughput phenotypic Platform (HTTP) to build the data set, and used RESNET 50 as the new feature extraction network instead of the original VGG 16 to optimize the Faster R-CNN model. According to the high time sequence and continuous images obtained by the HTTP, the number of male ears before and after tasseling stage in the experimental field was further monitored by using the improved model, and the tasseling stage was identified on this basis. The mean average precision and mean absolute error of the image data obtained by the method in the HTTP were 90.14% and 4.732 8 respectively. The mean average precision of the image data obtained from the UAV platform is 82.14%, and the mean absolute error is 9.694 8. The experimental results show that the model has better detection results on the HTTP than the UAV platform, and the model proposed in this paper has certain application value.