| 摘 要: 为解决在农田无人巡检过程中农作物目标智能识别的问题，将Cycle-GAN网络与Faster RCNN网络相结合构成了一种新的能支持少样本学习的网络模型，其中Cycle-GAN网络被用于提供数据增强。文中主要阐述了该网络的结构，并且对玉米、花生与豆子三种作物的种植地在不同的天气条件下进行了测试，对不同农作物的最优和最差识别率分别是96.53%和96.25%。实验证明，通过两者的结合能够提供更好更快的农作物识别和检测，新的少样本农作物目标识别模型具有较好的鲁棒性。 |
| 关键词: 少样本学习；农作物识别；数据增强；鲁棒性 |
|
中图分类号: TP311
文献标识码: A
|
| 基金项目: 国家级大学生创新训练项目(202011116001)；成都工业学院2020年青苗计划(2020QM085). |
|
| Research on an Improved Algorithm for Few-shot Crop Object Detection |
|
ZHANG Ruisen, WAN Xinghong, GAO Xin
|
(School of Automation and Electrical Engineering, Chengdu Technological University, Chengdu 611730, China)
3116738589@qq.com; 1720201171@qq.com; gaoandgao@126.com
|
| Abstract: In order to solve the problem of crop object intelligent detection in the process of unmanned farmland patrol inspection, this paper proposes a new network model that can support few-shot learning by combining Cycle-GAN network and Faster RCNN network. The Cycle-GAN network is used for data enhancement. The paper mainly explains the structure of the proposed network, and tests the planting areas of corn, peanuts and beans under different weather conditions. The best and worst detection rates for different crops under different weather conditions are 96.53% and 96.25%. Practice has proved that a combination of the two can provide better and faster crop identification and detection, and the new few-shot crop object detection model has better robustness. |
| Keywords: few-shot learning; crop detection; data enhancement; robustness |
