Semantic Segmentation of Weed and Crop with Partially Annotated Data for Automated Agriculture

Overuse of pesticides leads to severe environmental problems and increased cost of cultivation. As a reason precision agriculture is drawing research attention. The key idea is to use pesticides in a controlled manner targeting the weed. Therefore it is needed to detect weeds among crops accurately which could be accomplished by semantic segmentation. However, a key challenge with semantic segmentation in the needed training sets is the manual effort needed to label each pixel of each image. Towards this end, we explore two techniques namely marginal loss and background masking to perform semantic segmentation with partially annotated data. Two deep neural network models, U-Net and DeepLab V3+, are used as the backbone models in our evaluation with full annotation. We show that proposed methods achieve substantially accurate results with a very small amount of partially annotated data of real-world captured images used for training.