LAUSR

Abstract Mechanical weeding is an efficient weeding method, which is of considerable significance to the paddy field ecosystem. However, traditional mechanical weeding methods can cause seedling damages due to the bending phenomenon of the seedling lines. Introducing computer vision and control technology to traditional mechanical weeding methods can help the system diagnose the bending phenomenon and avoid crushing the seedlings. In this paper, we propose a deep-learning-based method of seedling line bending diagnosis and guidance line extraction. To prove the proposed method effective in the mechanical weeding system, we choose the Faster Region-based Convolutional Network (R-CNN) and Single Shot MultiBox Detector (SSD) as the representative models of the single-phase method and the two-phase method. With a novel dataset of rice seedling images established, we compare and analyze the confidence and real-time performance of the trained models. The experimental results show that the Faster R-CNN model is better in terms of accuracy, yet the SSD model has more advantages in the speed. Comprehensively considering the system requiring and model performances, the SSD model is a better choice in the automatic rice avoidance system.