A spontaneous keypoints connection algorithm for leafy plants skeletonization and phenotypes extraction

Zhen Wang, Xiangnan He, Yuting Wang, Chenxue Yang, Beilei Fan, Qingbo Zhou, Xian LiFrontiers in Plant Science 2025 · DOI

TL;DR

This paper introduces a novel, training-free algorithm for automatically extracting plant skeletons and phenotypes from images. By connecting spontaneously detected keypoints, it bypasses the need for manual annotation and lengthy training, making it highly scalable and adaptable for diverse plant structures.

What Problem Does This Paper Address?

Accurately measuring plant phenotypes is crucial for understanding growth and improving crop yields. Traditional methods often rely on manual measurements, which are time-consuming and prone to error, or sophisticated computer vision techniques that require extensive manual labeling, long training times, and predefined keypoints. This limits their scalability and adaptability to different plant species or even variations within a species. Existing deep learning approaches, while powerful, often struggle with the variability in leaf morphology, count, and occlusion found in complex plant structures. This paper aims to bridge this gap by developing a method that is both accurate and significantly less resource-intensive.

Methods

The core of this work is a "training-free and label-free" algorithm for skeletonizing leafy plants. The approach has two main branches, tailored for different leaf morphologies:

  • Random Leaf Morphology: For plants with irregular leaf shapes (like orchids), the algorithm starts by isolating leaf regions using color thresholding and morphological operations. It then generates random seed points within the leaf area. These points are connected iteratively by a local search that adapts its search direction based on the angle between consecutive points and a curvature minimization strategy. This process effectively traces the leaf skeleton.
  • Regular Leaf Morphology: For plants with more uniform leaf shapes, the algorithm fits the skeleton trajectory by minimizing curvature, ensuring a smooth and consistent representation.

The key innovation lies in the "spontaneous" detection and connection of keypoints. Instead of relying on predefined landmarks, the algorithm dynamically identifies and links points based on geometric constraints (angle difference, curvature) and local search strategies. This allows it to adapt to varying leaf shapes and numbers without prior knowledge.

Datasets

  • Orchid Dataset: A custom dataset of vertical and front-view images of orchids was collected using a multi-view automatic acquisition device. This dataset was used for developing and validating the algorithm, covering both random and regular leaf morphologies.
  • Maize Dataset: A publicly available dataset of individual maize plants was used to test the algorithm's generalization capabilities across different species.

Key Insights and Ideas

  • Training-Free Skeletonization: The most significant insight is the successful development of a skeletonization method that completely bypasses the need for model training and manual annotation. This drastically reduces the barrier to entry for plant phenotyping.
  • Adaptive Keypoint Connection: The algorithm's ability to adapt its keypoint connection strategy based on leaf morphology (random vs. regular) is crucial for its broad applicability.
  • Robustness to Variability: The method demonstrates remarkable robustness to variations in leaf count, density, and occlusion, which are common challenges in real-world plant imaging.
  • Cross-Species Generalization: The successful application to both orchids and maize highlights the algorithm's potential for wide-ranging use across different plant species.

Key Results

  • Orchid Performance: On orchid images, the algorithm achieved an average curvature fitting error of 0.12 and a leaf recall rate of 92%.
  • Phenotype Extraction: Five key orchid phenotypic parameters were accurately extracted from the generated skeletons.
  • Maize Performance: The method showed effective skeletonization on maize, demonstrating cross-species applicability.
  • Comparison with YOLOv7-SlimPose: The proposed method outperformed YOLOv7-SlimPose in terms of leaf recall rate (92% vs. 63%) and curvature error (0.12 vs. 0.59) on the test set, particularly in handling variable leaf counts and complex morphologies.
  • Computational Efficiency: The algorithm runs efficiently on a CPU, with average processing times of 3.556s for high-resolution images and requiring minimal memory.

Limitations and Future Work

  • Tip Detection Failure: The paper acknowledges that in some cases, particularly with very thin leaf tips, the algorithm might fail to detect keypoints at the outermost tip. This was observed to have a minor impact on most phenotypes but could be mitigated by reducing the marker diameter.
  • Maize Vertical View Misidentification: For maize in vertical view, the algorithm sometimes mistakenly identifies a single leaf as two separate ones. However, the authors note that this does not impact the extraction of key features like crown width, as the front view is primarily used for leaf count.
  • Future Work: The authors suggest that the method's effectiveness on both random and regular leafy plants makes it suitable for high-throughput plant phenotyping workflows. Further exploration into its application on an even wider range of plant species with complex architectures is implied.

Summary updated

Abstract

By eliminating manual labels and training, this approach reduces annotation effort and computational overhead while enabling precise geometric phenotype calculation from skeleton-based keypoints. Its effectiveness on both randomly distributed and regularly shaped leafy plants suggests suitability for high-throughput plant phenotyping workflows.
BibTeX
@inproceedings{wang2025spontaneous,
  author = {Zhen Wang and Xiangnan He and Yuting Wang and Chenxue Yang and Beilei Fan and Qingbo Zhou and Xian Li},
  title = {{A spontaneous keypoints connection algorithm for leafy plants skeletonization and phenotypes extraction}},
  booktitle = {Frontiers in Plant Science},
  year = {2025},
  doi = {10.3389/fpls.2025.1641255},
}
Figure