Interpretable Leaf Vein Analysis

An explainable image-processing pipeline for extracting and measuring leaf vein networks.

Python
OpenCV-style image processing
PCA
K-Means
Random Forest
Jupyter

Conceptual illustration created for this case study — not a screenshot of the application.

Overview and context

An interpretable computer-vision pipeline that extracts, measures, and analyses leaf vein networks from grayscale images, combining classical image processing, handcrafted structural features, dimensionality reduction, clustering, and machine-learning validation. Completed August 2026 as part of my MSc.

Problem

Faint secondary veins, uneven lighting, noise, and changing vein thickness make leaf vein extraction unreliable — and deep networks can hide why they succeed or fail.

My role and contribution

Co-author of the project with Annur Anika; the project write-up credits both authors jointly.

Technical approach

Standardize images to 512 × 512 and normalize intensity.
Compare Gaussian, median, and bilateral filtering for noise reduction versus structure preservation.
Enhance local contrast with CLAHE, unsharp masking, and top-hat transformation.
Segment with global, Otsu, and adaptive thresholding plus Canny edges; refine with morphology and connected components.
Skeletonize the network, detect endpoints and junctions, extract handcrafted features, then apply PCA, K-Means clustering, and Random Forest validation.

Evidence and outcomes

A complete, inspectable pipeline applied to 900 CHAMBASA leaf images.