AI-powered Weed Identification in cultivated fields: an Object Detection approach

The goal of this thesis is to analyze the use of object detection in solving weed extirpation issues in agriculture. First, I will give an overview of different weed detection systems and analyze the advantages and disadvantages of the state-of-the-art object detection algorithms. In particular, I will focus on YOLO as a single-stage detector and Faster R-CNN as a two-stage detector. To this end, I have created a dataset of images submitted by farmers and labeled them with crops, weeds, weed roots, and rows of crops using the LabelImg tool. I run several experiments on this dataset and compare the two detectors in terms of evaluation metrics. In addition, it is fundamental to also focus on the inference time of the models since the system is designed to be used in a real-time scenario. Therefore, I test OpenVINO and ONNX, two toolkits designed for optimizing and delivering AI inference, to speed up the inference phase. As a result, YOLO proves to be the most suitable detector for the purposes of this project, while ONNX and OpenVINO significantly speed up the model at the expense of slight accuracy losses.