1State Key Laboratory of Rice Biology and Breeding, China National Rice Research Institute, Chinese Academy of Agricultural Sciences, Hangzhou, 310006, Zhejiang, China
2School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou, 310018, Zhejiang, China
3College of Agriculture, Yangtze University, Jingzhou, 434025, Hubei, China
4These authors contributed equally to this work.
| Received 07 Jun 2024 |
Accepted 23 Dec 2024 |
Published 28 Feb 2025 |
Location-based methods for counting rice panicles have often been underestimated, primarily due to their perceived inferior performance when compared to detection-based techniques. However, we argue that the potential of these location-based methods has not been fully realized, largely owing to the limitations of existing model architectures. In response to this challenge, we introduce LKNet, an innovative model developed on the foundation of the location-based framework P2Pnet. To enhance the performance of panicle counting across diverse types and growth stages, we implemented several key strategies. Firstly, we reconstructed the localization loss function as a predictive probability distribution to reduce the influence of manual labeling. Additionally, we dynamically adapted the receptive field to better accommodate different panicle types through the use of large kernel convolutional blocks. We evaluated LKNet on several publicly available counting task datasets and achieved state-of-the-art performance on the Diverse Rice Panicle Detection dataset. Furthermore, we employed a rice panicle dataset collected at an altitude of 7 m, which includes various panicle types and growth stages for model training and evaluation. The results showed that LKNet effectively accommodates variations in panicle morphology, with R2 values ranging from 0.903 to 0.989. These findings highlight LKNet's potential to enhance precision in panicle counting in rice breeding programs.
