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

Area of Science:

• Plant Science
• Computer Vision
• Computational Biology

Background:

• Leaf phenotypes are crucial for assessing plant growth.
• Current deep learning methods for leaf skeletonization demand extensive manual labeling, long training times, and predefined keypoints, hindering scalability.
• There is a need for efficient and scalable automated methods for leaf skeletonization and phenotype extraction.

Purpose of the Study:

• To develop a training-free and label-free approach for generating leaf skeletons from images of leafy plants.
• To enable precise calculation of geometric plant phenotypes from the generated leaf skeletons.
• To reduce the annotation effort and computational overhead associated with traditional leaf skeletonization methods.

Main Methods:

• The approach involves random seed-point generation and adaptive keypoint connection.
• It utilizes an angle difference threshold for plants with random leaf morphology and curvature minimization for regular leaf morphology.
• Validation was performed on orchid and maize image datasets.

Main Results:

• The method achieved an average curvature fitting error of 0.12 and 92% leaf recall on orchid images.
• Five phenotypic parameters were accurately extracted from orchid leaf skeletons.
• Effective skeletonization was demonstrated on maize, indicating cross-species applicability.

Conclusions:

• The training-free and label-free approach significantly reduces annotation and computational requirements for leaf skeletonization.
• The method enables precise geometric phenotype calculation, suitable for high-throughput plant phenotyping.
• Its effectiveness across diverse leaf morphologies (random and regular) highlights its broad applicability in plant science research.