Abstract: Weed is an important factor leading to crop yield loss. Aiming at the problems of low accuracy and small range of weed identification in the natural environment of the field, an identification algorithm of field weeds based on improved Faster R-CNN and transfer learning is proposed. Firstly, weed images of different periods and angles in multiple scenes were collected, and the data set expended by rotation, clipping, color adjustment, etc. Then, based on the original Faster R-CNN, the improved dualthreshold nonmaximum suppression algorithm was used to find the bounding boxes with high confidence. Finally, AlexNet, GoogleNet, VGG16 and ResNet50 were used as the region proposal network of the model, and the optimal model parameters were transferred to the filed weed identification task. Through testing and verification on multispecies dataset and small sample dataset, the experimental results showed that our method could achieve precision of 9658%, recall of 94.82% and F1-score of 95.06%, and it has a wider identification range than current mainstream methods while maintaining a higher identification accuracy.

Key words: weeds identification, Faster R-CNN, transferring learning, dualthreshold nonmaximum suppression algorithm

摘要： 杂草是导致农作物减产不保量的重要因素，针对田间自然环境下杂草识别精度低和识别范围局限的问题，提出一种基于改进Faster R-CNN与迁移学习的农田杂草识别算法。首先，采集多场景下不同时段不同角度的杂草图片，通过旋转、裁剪和调节色彩等方式扩充数据集；然后，在原始Faster R-CNN网络的基础上利用改进的双阈值非极大抑制算法（NonMaximum Suppression，NMS）查找置信度较高的边界框；最后，将AlexNet、GoogleNet、VGG16和ResNet50等作为模型的区域建议网络，并将其最优模型参数迁移至农田杂草识别任务中。通过在多样本数据集和少量物种样本数据集上进行测试验证，试验结果表明，算法可以实现96.58%的精确率、94.82%的召回率和95.06%的F1-score，相比当前主流算法在保持识别精度较高的基础上，具有更广的识别范围。

关键词: 杂草识别, Faster R-CNN, 迁移学习, 双阈值非极大抑制算法