The rate of missing rice seedlings is key information in the early stages of rice cultivation, which affects farmers' decision-making process, such as replanting. Due to the low efficiency and poor accuracy of traditional manual seedling counting methods, an automated method for accurately detecting missing rice seedlings is needed. Therefore, this study proposed a real-time missing rice seedling counting method based on object detection and tracking-by-detection algorithm, including improved YOLOv5s and ByteTrack. The study focused on two aspects: improving the detection accuracy of tiny and overlapping seedlings and constructing lightweight models. The detection accuracy was improved by adjusting the structure of the detection head and introducing Transformer encoders. Meanwhile, the methods of channel pruning and compact network construction were used for developing lightweight models. The experimental results showed that the improved YOLOv5s network had a mAP@0.5:0.95 of 72.3% and an FPS of 71.6f/s, significantly higher than the original YOLOv5s model (the mAP@0.5:0.95 was 71.8% and the FPS was 56.1f/s). The accuracy of the missing seedlings counting method was 93.2%, and the counting time was five times less than the manual counting time. Therefore, the proposed method is considered both practical and efficient in detecting and counting missing seedlings, providing a new solution for real-time counting of missing rice seedlings in paddy fields.
Read full abstract