Grid Cell Detection of Dandelion Weed Centers via Convolutional Neural Network

Abstract

This paper presents a novel method for detecting dandelion weed (Taraxacum officinale) plant centers in perennial ryegrass using partial information gathered only from plant leaves. A primitive region proposal method generates proposals from original birds-eye view images of whole dandelion weeds in grass. The proposals containing dandelion weed leaves are taken and plant centers are labeled with a point based on the novel concept of the most “prominent” leaf. The samples are divided into a grid of cells and the cell containing the labeled point is considered the truth cell. A radial map and its inverse are generated based on the spatial location of the cells w.r.t. the truth cell. A fully convolutional network is trained to detect the positive truth cell using novel loss functions based on these maps. Using a relatively small dataset, the loss functions with the terms that compute regression loss on the maps yield significantly better model performance than those without. In addition, some errors are simply the result of the center of an alternate “prominent” leaf being automatically detected. Further, the comparison results with segmentation models reveal some advantages in detecting only plant centers as opposed to training computationally costly inference models.