2020 Volume 82 Issue 1 Pages 47-56
In order to navigate a robotic combine harvester, we have developed a rice field image segmentation methods based on deep learning architectures. Based on several classic convolutional neural network models, eight different segmentation models were constructed. After training, the speed and accuracy of all models processing the same dataset were analyzed and compared. The results indicated that the entire group of models were able to effectively perform the segmentation of the harvested areas, unharvested areas, ridge areas, and background areas in the images when the unmanned rice combine harvester was in operation. The pixel accuracy, class mean accuracy and mean IU (intersection over union) of the best model were 95.17 %, 86.08 %, and 80.07 %, respectively.
