基于深度学习的温室大棚小番茄果实计数方法

Counting method of cherry tomato fruits in greenhouses based on deep learning

目的:以小番茄果实簇为目标,将果实计数任务转化为果实簇的检测和分类两个子任务,有效避免了果实间的重叠、遮挡问题,并大大减少了标注工作量,进而提出两种基于深度学习的番茄果实计数方法。方法:方法1采用YOLOv3网络直接对果实簇进行检测和分类,实现果实计数;方法2首先采用YOLOv3网络检测果实簇,再通过ResNet50网络根据果实数不同对果实簇进行分类,最后统计总的果实数。结果:实验表明,两种方法均能实现小番茄果实的计数,其中方法2采用两种网络分别进行果实簇的检测和分类,计数精度高于方法1,在测试样本上的预测果实数与实际果实数之间的均方根误差(Root mean square error,RMSE)为6.37个,平均绝对百分误差率(Mean absolute percentage error,MAPE)为13.89%。结论:本文所提以果实簇为目标的计数方法可有效实现小番茄果实的计数,对簇状果实的产量估计研究具有一定参考价值。

Aims:In this paper,the fruit counting task was transformed into two subtasks:detection and classification of fruit clusters,which effectively avoided the problem of serious overlapping between single fruits and greatly reduced the labeling workload.Two kinds of tomato fruit counting methods based on deep learning were proposed.Methods:Method 1 used YOLOv3 network to detect and classify fruit clusters directly and realized fruit counting.Method 2 first applied YOLOv3 network to detect fruit clusters and then utilized ResNet50 network to classify fruit clusters according to fruit numbers and counted the total number of fruits.Results:The experimental results showed that the two kinds of methods achieved cheery fruit counting.Method 2 used two networks to detect and classify fruit clusters respectively;and the counting accuracy was higher than Method 1.The root mean square error(RMSE)and the mean absolute percentage error(MAPE)of the test sample were 6.37 and 13.89%,respectively.Conclusions:The counting method with fruit clusters as the target proposed in this paper can effectively count cherry tomato fruits,which has certain reference value for the yield estimation research of cluster fruits.