基于Faster R-CNN网络的茶叶嫩芽检测

为有效识别茶叶嫩芽提高机械采摘精度、规划采摘路线以避免伤害茶树，针对传统目标检测算法在复杂背景下检测精度低、鲁棒性差、速度慢等问题，探索了基于Faster R-CNN目标检测算法在复杂背景下茶叶嫩芽检测方面的应用。首先对采集图像分别进行等分裁切、标签制作、数据增强等处理，制作VOC2007数据集；其次在计算机上搭建深度学习环境，调整参数进行网络模型训练；最后对已训练模型进行测试，评价已训练模型的性能，并同时考虑了Faster R-CNN模型对于嫩芽类型(单芽和一芽一叶/二叶)的检测精度。结果表明，当不区分茶叶嫩芽类型时，平均准确度(AP)为54%,均方根误差(RMSE)为3.32;当区分茶叶嫩芽类型时，单芽和一芽一叶/二叶的AP为22%和75%,RMSE为2.84;另外剔除单芽后，一芽一叶/二叶的AP为76%,RMSE为2.19。通过对比基于颜色特征和阈值分割的茶叶嫩芽识别算法(传统目标检测算法),表明深度学习目标检测算法在检测精度和速度上明显优于传统目标检测算法(RMSE为5.47),可以较好地识别复杂背景下的茶叶嫩芽。

Tea Bud Detection Based on Faster R-CNN Network

Effective detection of tea buds is an important prerequisite for improving the precision of mechanical picking and planning the picking route to avoid harming tea plants. Considering the problems of low detection accuracy, poor robustness and slow speed of traditional target detection algorithm in complex background, Faster R-CNN was applied to recognize tea bud in complex background. Firstly, collected pictures were processed by equal cutting, label making and data enhancement to make VOC2007 dataset. The deep learning model on detecting tea bud types (single bud and one bud with one leaf/two leaves) was trained after setting up the environment and adjusting the model parameters, and the trained model was evaluated. The results showed that the average precision (AP) was 54%, and the root mean square error (RMSE) were 3.32 when the tea bud type was not distinguished. When distinguishing tea bud types, the AP of single bud and one bud with one leaf/two leaves were 22% and 75%, with RMSE of 2.84. When single bud was removed, the AP of one bud with one leaf/two leaves was 76%, with RMSE of 2.19. Compared with tea bud detection algorithm based on excess green and image binarization (traditional target detection algorithm), the deep learning target detection algorithm was superior to traditional target detection algorithm, with RMSE of 5.47, in accuracy and speed, especially under complex background. Deep learning algorithm demonstrated an important application prospect in realizing tea bud detection and automatic picking in intelligent tea garden image real-time detection system.