This paper proposes a method for training deep convolutional neural networks (DCNNs) as a detector of plant stems partly hidden by leaves. The detector is assumed to be used by an agricultural robot that treats near trees in dense fruit vegitable fields, where the stems of trees are hidden not only by leaves, but also by the totally green confusing background. We tackle this difficult problem by training DCNNs with datasets that are compose of realistic computer graphic images (CG images) and annotated images of foreground leaves and stems. Physically-based Rendering (PBR) and Image-based Learning (IBL) are utilized for creating the CG images that make DCNNs distinguish foreground trees from the background. We have trained two DCNNs with the CG images and created a main stem detector with the DCNNs. In the experiments, the detector has distinguished near trees from the confusing background and has found a main stem partly hidden by leaves.
