Data Augmentation Using GANs for Crop/Weed Segmentation in Precision Farming

Farming robots need a fast and robust image segmentation module to apply targeted treatments, which require the ability to distinguish, in real time, between crop and weeds. Existing solutions make use of visual classifiers that are trained on large annotated datasets. However, generating large datasets with pixel-wise annotations is an extremely time-consuming task. In this work, we tackle the crop/weed segmentation problem by using a synthetic image generation method to augment the training dataset without the need of manually labelling the images. The proposed approach consists in training a Generative Adversarial Network (GAN), which can automatically generate realistic agricultural scenes. As a difference with respect to common GAN approaches, where the network learns how to reproduce an entire scene, we generate only instances of the objects of interest in the scene, namely crops. This allows to build a generative model that is more compact and easier to train. The generated objects are then placed into real images of agricultural datasets, thus creating new images that can be used for training. To evaluate the performance of the proposed approach, quantitative experiments have been carried out using different segmentation network architectures, showing that our method well generalizes across multiple architectures.