Weed segmentation from grayscale tobacco seedling images

Abstract

Manual weed extraction from young seedlings is a hard manual labour process. It has to be continuously performed to increase the yield per land unit of any agricultural product. Precise segmentation of plant images is an important step towards creating a camera sensor for weed detection. In this paper we present a machine learning approach for segmenting weed parts from images. A dataset has been generated using bumblebee camera under various light conditions and subsequently training and test patches were extracted. We have generated various texture-based descriptors and used different classification algorithms aiming at correctly recognizing weed patches. The results show that in a case when the images are grayscale, the light conditions are varying, and the distance of the camera to the weeds is not constant machine learning algorithms perform poorly.