The Study of Light-weight YOLOv4 Model for Rice Seedling and Counting

Abstract

Rice is a very important agricultural product, especially in Asia country such as Japan, Thailand, etc. It is a daily essential food for many people. To better monitor the rice yield, it is necessary to pay attention to the rice seedling stage. In the past, many scholars have used image processing technologies to complete the counting of rice seedlings. However, it is common that the color of rice changing accordance with the changing weather, which may cause the counting error if using the traditional image processing method. It is also possible that there are weeds or other non-rice obstructions that confuse the image recognition and create counting errors. In the past, not many scholars used object detection technology to locate rice seedlings, however, it is important to identify the rice object for counting. Hence, this research applies the YOLO model to explore the object detection technology to complete the positioning and counting of rice seedlings. To improve the model performance, the YOLOv4 architecture was deeply studied and adjusted, to reduce the training process and training time, thereby achieving the purpose of a lightweight model, we named it as YOLO4-L1. In this study, we established a system for automatic positioning of object detection and calculation of rice seedlings. Comparisons among our proposed YOLO4-L1 model with YOLOv3-tiny, YOLOv4-tiny, YOLOv3, and YOLOv4 are conducted. Our experimental results have shown that our proposed YOLO4-L1 model can reduce 2.45hr for training time with similar counting result when comparing with YOLOv4 model.