PFL-Net: a deep learning network for maize leaf point cloud completion
WU Zhangbin, Weiliang Wen, Anran Song, Yandong Wu, Teng Miao, Xinyu Guo
National Engineering Research Center for Information Technology in Agriculture Shenyang Agricultural University Beijing Academy of Agriculture and Forestry Sciences
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Maize leaves represent a pivotal component in the architecture of the plant and the process of photosynthesis. Leaf point clouds have been shown to accurately reconstruct the three-dimensional morphology of real leaves, thus rendering them a vital tool for research in the fields of maize breeding and cultivation. However, the maize leaf point clouds that are the result of this process often contain a certain degree of data loss. This is due to a number of factors, including obstruction by plant organs, surface reflections from the leaf material, and the ambient lighting conditions. In addition, the method of data acquisition has a significant impact on subsequent leaf phenotyping and visualization analyses. Existing point cloud completion methods frequently encounter difficulties in achieving satisfactory results when applied to the elongated, narrow configuration of maize leaf point clouds. In order to address this issue, the present study proposes a deep learning-based maize leaf point cloud completion network, which is termed Point Fractal Leaf Network (PFL-Net). Initially, four distinct methodologies were employed to collect leaf point cloud data, with the objective of establishing a high-quality maize leaf point cloud dataset. PFL-Net is built upon Point Fractal Network (PF-Net) by introducing a Graph Block (GB) module to better capture edge features in point clouds. Moreover, we extend the original Combined Multi-Layer Perception (CMLP) encoder into an Average Combined Multi-Layer Perception (ACMLP) module, directing the model's attention toward fine-grained geometric details. PFL-Net outperforms PF-Net across all five evaluation metrics, reducing Earth Mover's Distance (EMD) by 0.013 and improving F1-score by 0.019. Notably, biomass prediction based on PFL-Net-reconstructed leaves achieves an R-squared ( R ²) value identical to that obtained from complete leaf scans. These results position PFL-Net as a key enabler for 3D phenotyping and plant architecture studies in maize breeding and cultivation.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AI3D Shape Modeling and Analysis
Smart Agriculture and AI · Greenhouse Technology and Climate Control