Abstract

Current research in agricultural weeding automation attempts to develop accurate methods of distinguishing between crop and weed. Consequently, the use of computer vision has become a cornerstone in these endeavours. Some recent methods employ pattern recognition techniques that involve hierarchical feature groupings. The application generally applies some form of machine learning. Furthermore, using convolutional neural networks (CNN), many techniques implement complex architectures that not only classify but also detect and locate objects. These detection problems generally involve datasets taken under artificial or controlled lighting conditions where foreground elements (i.e. weed and crop) are easily distinguishable from the background (usually soil) by virtue of their distinct hue and textures. Plant overlap is generally limited to being between foreground elements. The research in this thesis addresses the challenges overlooked by agricultural weeding by focusing on weeding in lawn grass with two distinct approaches. First, a pattern recognition methodology is developed to distinguish dandelion weed centers from grass using the morphological attributes of binary (black-and-white) regions. This method is tested in lab settings with both artificial weeds and grass. However, practical limitations include a fragile performance in real-world applications in the field and a heavy reliance on parameter calibration. Next, a machine-learning approach is developed to address the shortcomings of the prior approach as well as to deal with the challenges specific to weeding in a domestic setting. A five-step process involving CNN structures proves successful at accurately detecting dandelion weeds within grass and other lawn vegetation. Extensive tests have been carried out on a wide array of real work images and the results demonstrate that the developed algorithm can detect and recognize dandelions in the grass within a reasonable range of natural lighting conditions.