Semi-supervised semantic segmentation for grape bunch identification in natural images

In precision farming, the actual implementation of plant monitoring requires low-cost devices, which often return data of poor quality. This increases the complexity of the processing steps which need advanced tools, such as deep learning methods. In this work, three deep architectures, namely the DeepLab, the HRNet, and the U-Net, for the semantic segmentation of natural images of a vineyard have been compared, and a semi-supervised PseudoLabeling technique is proposed to take advantage of non-annotated images. In these experiments, the DeepLab architecture best-performed with a mean segmentation accuracy of the bunch class of 84.37%, improving the previously existing models by 3.79%, whereas PseudoLabeling further boosted its performance by an additional 1.78%.