Mask R-CNNを用いたニンニク圃場における雑草検出の最適化と実装

抄録

Robotic weeding manipulation must remove weeds without damaging crops. To achieve this, the robot must be able to detect individual weeds in the image of fields where crops and weeds are mixed and operate the weeding device on those coordinates. We develop a system to detect weeds using Mask-RCNN algorithm from the image of ridges captured by Intel Realsense, and evaluate the effect of the annotation shape of the training data used in this algorithm. We compare the performances of square and polygonal segmentations. Results of comparing based on a single image, accuracy of polygonal segmentation is better than that of square segmentation. In addition, polygonal segmentation has a segmentation mask which each detected weeds along the contours.