Development of Deep Learning-Based Laser Weeding Robot

Agriculture is an important sector that contributes significantly to Vietnam's GDP. The use of pesticides and herbicides, however, raises great concern due to their harm to the environment and human health. The government is taking action to eliminate the use of these chemicals, meanwhile, alternative solutions are needed. The aim of this paper is to develop an effective solution for weed control using a computer vision-based laser weed-killing robot. To ensure efficiency and accuracy, computer vision models must be evaluated based on different metrics such as performance, speed, and computational cost represented by mean average precision, speed of inferences (in fps), and the size of the model in FLOPs and params. The study focuses on deploying the state-of-the-art object detection model YOLOv8 on a dataset of 9,000 labeled images taken from fields in Pakistan. YOLOv8 outperforms YOLOv5 in a previous study with an accuracy of 0.9043@IOU 0.5, particularly suited for hardware efficient integration due to its 'n' and ‘s’ versions. Both models in “n” version demonstrated similar real-time frame processing times, operating at an FPS rate of approximately 62.5ms using the same computer configuration. The results show that YOLOv8n is a fitting choice for the laser weedkilling robot due to its superior weed detection capabilities.