Generating a Dataset for Semantic Segmentation of Vine Trunks in Vineyards Using Semi-Supervised Learning and Object Detection

摘要

This article describes an experimentally tested approach using semi-supervised learning for generating new datasets for semantic segmentation of vine trunks with very little human-annotated data, resulting in significant savings in time and resources. The creation of such datasets is a crucial step towards the development of autonomous robots for vineyard maintenance. In order for a mobile robot platform to perform a vineyard maintenance task, such as suckering, a semantically segmented view of the vine trunks is required. The robot must recognize the shape and position of the vine trunks and adapt its movements and actions accordingly. Starting with vine trunk recognition and ending with semi-supervised training for semantic segmentation, we have shown that the need for human annotation, which is usually a time-consuming and expensive process, can be significantly reduced if a dataset for object (vine trunk) detection is available. In this study, we generated about 35,000 images with semantic segmentation of vine trunks using only 300 images annotated by a human. This method eliminates about 99% of the time that would be required to manually annotate the entire dataset. Based on the evaluated dataset, we compared different semantic segmentation model architectures to determine the most suitable one for applications with mobile robots. A balance between accuracy, speed, and memory requirements was determined. The model with the best balance achieved a validation accuracy of 81% and a processing time of only 5 ms. The results of this work, obtained during experiments in a vineyard on karst, show the potential of intelligent annotation of data, reducing the time required for labeling and thus paving the way for further innovations in machine learning.