Comparison of One-Stage Object Detection Models for Weed Detection in Mulched Onions

Abstract

Deep learning-based computer vision enabled farming robots to detect and control weeds in the field accurately. This study compared the performance of Scaled-YOLOv4-CSP, YOLOv5s, and SSD Mobilenet V2 for image-based weed detection in mulched onions. The study showed that YOLOv5s is more suitable for the purpose. At 0.915 $\mathrm{m}\mathrm{A}\mathrm{P}^{0.5}$, it ties with Scaled-YOLOv4-CSP at first place on weed detection performance. Meanwhile, YOLOv5s consumed significantly fewer resources during training and implementation, giving it an advantage in real-time weed detection. Its mean inference time of 7.72 milliseconds is also less than half of the other two models. Lastly, the study demonstrated that increasing the number of samples with a more balanced class distribution by upsizing the dataset through data augmentation would improve the overall performance of the object detection model.