Revealing hierarchical structure of leaf venations via diffusion-refined label-efficient segmentation: Dataset and...

Area of Science:

• Plant Science
• Computational Biology
• Image Analysis

Background:

• Understanding leaf vein hierarchy is vital for plant physiology, evolution, and ecology.
• Existing data-driven segmentation methods lack specialized datasets for hierarchical leaf vein analysis.

Purpose of the Study:

• Introduce the Hierarchical Leaf Vein Segmentation (HALVS) dataset, the first of its kind.
• Propose a novel, label-efficient hierarchical segmentation framework for leaf veins.

Main Methods:

• Developed the HALVS dataset with 5057 images and extensive annotations across three species.
• Proposed a framework combining Partially Supervised Semantic Segmentation (PSSS) and Denoising Diffusion Label Refinement (DDLR).
• Pioneered cross-species learning for leaf vein segmentation.

Main Results:

• The HALVS dataset provides high-definition images and detailed annotations for hierarchical vein analysis.
• The PSSS-DDLR framework significantly improves tertiary vein integrity and connectivity with reduced annotation costs.
• Cross-species learning demonstrated robust performance, even on challenging species.

Conclusions:

• The proposed framework offers an effective solution for segmenting complex hierarchical leaf vein patterns.
• This work advances agricultural research by enabling more accurate and efficient leaf vein analysis.
• The HALVS dataset and framework address a critical need in plant science research.