Introduction
In botany and agriculture, the "plumbing" of a leaf—its venation network—serves as a critical indicator of plant health, photosynthetic efficiency, and drought resistance. While computer vision has previously been employed to segment these networks, most existing models treat all veins as a single class. In reality, leaf veins are hierarchical: primary (1°) veins are the thick central "highways," secondary (2°) veins branch off them, and tertiary (3°) veins form a fine, dense mesh.
Accurately segmenting this hierarchy is challenging due to two main factors: the lack of public datasets distinguishing these levels and the extreme cost of annotation. Labeling intricate 3° veins is a massive human effort; in this study, fully labeling just 150 images required four experts nearly 84 person-days. This research introduces a new dataset and a label-efficient training method to make hierarchical segmentation feasible.
The HALVS Dataset
To address data scarcity, the authors present the HierArchical Leaf Vein Segmentation (HALVS) dataset. It consists of 5,057 high-resolution images (600 DPI) of soybean, sweet cherry, and London planetree leaves. Unlike previous datasets that often use chemically cleared leaves, HALVS uses untreated leaves captured via a flat-bed scanner in transmission mode, providing high contrast suitable for real-world agricultural applications.
The dataset includes 150 "fully labeled" images (1°, 2°, and 3° veins), while the remainder are either unlabeled or "partially labeled" (containing only 1° and 2° vein annotations). This structure reflects the practical reality of botanical data collection, where the finest details are the hardest to document.
Methodology: Partially Supervised Semantic Segmentation (PSSS)
The core technical contribution is the Partially Supervised Semantic Segmentation (PSSS) module. This pluggable module can be integrated into existing semi-supervised learning (SSL) frameworks like FixMatch or UniMatch. It is specifically designed to handle images where 3° veins are missing from the ground truth.
The PSSS module utilizes a teacher-student architecture to categorize pixels into three sets: known (1° or 2°), confident pseudo-labels (background or 3°), and uncertain pixels. A key innovation is the Class-Specific Exclusion Loss. By leveraging the biological fact that vein types are mutually exclusive, the model can infer the presence of 3° veins by excluding other classes. If the model reliably predicts a pixel is not background, 1°, or 2°, it is forced to conclude it must be a 3° vein, effectively "filling in the blanks" of incomplete human annotations.
Key Results and Insights
The PSSS method demonstrates significant performance gains and efficiency:
- Performance Boost: Integrating the PSSS module into the UniMatch baseline improved the Intersection-over-Union (IoU) for 3° veins by 11% and the overall mean IoU (mIoU) by 6.82%.
- Label Efficiency: The method achieved results competitive with an "Oracle" model (trained on fully supervised data) while requiring significantly less manual labor.
- Cross-Species Transfer: The study found that training on simpler leaf structures (like soybean) and transferring to more complex ones (like planetree) is a viable strategy when the PSSS module handles the target species' unlabeled data.
Limitations and Future Directions
Despite these advancements, hierarchical segmentation remains a challenging task. Performance gaps persist when moving between species with vastly different morphologies. Additionally, the finest tertiary veins, often only a few pixels wide, remain susceptible to noise even in high-resolution scans. Future work may involve exploring more advanced cross-species transfer learning and incorporating botanical structural priors—such as branching patterns—directly into neural network architectures to further refine segmentation accuracy.
Summary updated