Artificial intelligence-based semi-supervised crop and weed semantic segmentation

Abstract

Accurate segmentation of crop and weed by farming robot camera can increase crop production and reduce unnecessary herbicide, which is a fundamental task in the field of sustainable and precision agriculture. However, obtaining the pixel-wise annotation of training data manually is expensive. As a solution to address this limitation, semi-supervised learning leverages a small amount of labeled data and a large amount of unlabeled data for learning. In this context, we propose a network based on vector quantization and prototype loss for semi-supervised crop and weed semantic segmentation (VQP-Net). VQP-Net achieves a strong performance in terms of consistency regularization through the implementation of a vector quantization module and prototype loss, and is capable of extracting discriminative features of crops and weeds, which are often indistinguishable. We conducted experiments using the proposed method with three open datasets: BoniRob, crop/weed field image, and rice seedling and weed datasets. The crop and weed segmentation accuracies based on mean intersection over union (mIOU) for the three datasets were 0.8643, 0.8329, and 0.7623, respectively, demonstrating that this method outperformed the state-of-the-art methods.