Precision Agriculture Science and Technology ISSN:2672-0086(Print) 2713-5632(Online) 정밀농업과학기술

In this study, a weeding robot with real-time crop recognition and weed control techniques was developed. To collect real-time images while the robot was moving, a self-developed image acquisition device using a mono camera and a Jetson board was installed on the weeding robot. Data analysis and deep learning were performed using Python, PyTorch and Torchvision. A weakly supervised learning approach was used for bean detection and a class activation map method was used for deep learning training. The layout was designed to achieve variable automatic weeding, where the weed control unit would rotate at a certain angle to avoid damaging the plant when the bean plant was detected in the real-time image captured by the robot's camera. The training results of the deep learning model showed that the accuracy and loss of the model converged with repeated training, and the recall rate was 0.85. The recognition accuracy of the bean plant area was confirmed to be 95.2%. The automation of the weed control unit resulted in successful avoidance of the bean plant when tested in five trials, with a success rate of 98.7%. The results confirmed that this approach is suitable for practical application in actual weed control operations.