FACNet: A high-precision pumpkin seedling point cloud organ segmentation method

Abstract

Accurately segmenting plant organs in pumpkin seedling point clouds is crucial for automated plant phenotyping and is essential for improving cultivation efficiency and optimizing breeding strategies. The segmentation of pumpkin seedling organs in point clouds presents challenges such as leaf overlap, indistinct boundaries between stems and leaves, and the morphological diversity of leaves and stems. To address these issues, we propose a high-precision pumpkin seedling point cloud organ segmentation network and construct, for the first time, a labeled dataset for pumpkin seedling point cloud organ segmentation. Firstly, to tackle the difficulties caused by leaf overlap and ambiguous stem-leaf boundaries, we present the Fused Bilinear Feature Extractor (FBFE). This method utilizes bilinear operations to combine local and global features, precisely capturing subtle feature differences at overlapping leaves and stem-leaf junctions. Secondly, to address the challenge of morphological diversity between leaves and stems, we introduce the Adaptive Multi-Scale Feature Fusion Module (AMSF). This module automatically adjusts feature fusion strategies across different scales, effectively integrating information from various levels and enhancing the model’s ability to handle morphological diversity and capture fine details. Finally, we propose the Chebyshev Particle Snow Ablation Optimizer (CPSAO) to optimize the learning rate, improving the model’s convergence speed and segmentation accuracy. Experimental results show that FACNet achieves 95.06 % mIoU, 96.87 % mPrec, 98.02 % mRec, and 97.44 % mF1 on the pumpkin seedling point cloud segmentation dataset. Compared to popular point cloud segmentation models, FACNet offers superior precision and stability in segmenting organs from pumpkin seedling point clouds.

Plant phenotyping relevance

カボチャ幼苗の点群から器官を抽出するセグメンテーション手法を開発し、ラベル付きデータセットも構築しており、植物表現型取得の方法が中心的です。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.