[Paper Review] Progressive Tree-like Curvilinear Structure Reconstruction with Structured Ranking Learning and Graph Algorithm

Seong-Gyun Jeong, Yuliya Tarabalka+2 more|arXiv (Cornell University)|Dec 8, 2016

Automated Road and Building Extraction28 references3 citations

TL;DR

This paper proposes a novel graph-based method for reconstructing tree-like curvilinear structures using structured ranking learning and geodesic path optimization. By learning orientation-aware feature rankings via Structured SVM and progressively building a topologically accurate structure from minimal pixels, the approach achieves state-of-the-art F1 scores and significantly reduces pixel usage across diverse datasets like retinal vessels, road networks, and cracks.

ABSTRACT

We propose a novel tree-like curvilinear structure reconstruction algorithm based on supervised learning and graph theory. In this work we analyze image patches to obtain the local major orientations and the rankings that correspond to the curvilinear structure. To extract local curvilinear features, we compute oriented gradient information using steerable filters. We then employ Structured Support Vector Machine for ordinal regression of the input image patches, where the ordering is determined by shape similarity to latent curvilinear structure. Finally, we progressively reconstruct the curvilinear structure by looking for geodesic paths connecting remote vertices in the graph built on the structured output rankings. Experimental results show that the proposed algorithm faithfully provides topological features of the curvilinear structures using minimal pixels for various datasets.

Motivation & Objective

• To address the challenge of reconstructing topologically accurate curvilinear structures in images with low contrast and complex geometries.
• To overcome limitations of threshold-based segmentation and centerline methods that fail to preserve junctions and connectivity.
• To learn structured rankings of image patches that reflect plausibility to latent curvilinear patterns using supervised learning.
• To reconstruct curvilinear structures progressively by identifying geodesic paths on a graph built from ranking scores.
• To minimize pixel usage while preserving topological fidelity across diverse datasets.

Proposed method

• Compute oriented gradient features using steerable filters to extract local curvilinear patterns from image patches.
• Train a Structured Support Vector Machine (SSVM) for ordinal regression to rank patches based on shape similarity to latent curvilinear structures.
• Align each patch to a baseline orientation (horizontal) using local major orientation to ensure rotational invariance in feature representation.
• Construct a graph where nodes correspond to ranked image patches and edges represent spatial connectivity.
• Perform progressive reconstruction by iteratively identifying and extending geodesic paths between remote vertices in the graph.
• Use a topological importance ordering to prioritize the addition of major branches and fine details in the reconstruction sequence.

Experimental results

Research questions

• RQ1 Can structured ranking learning effectively capture the plausibility of image patches to latent curvilinear structures?
• RQ2 Can geodesic path computation on a ranking-based graph reconstruct topologically accurate curvilinear structures with minimal pixel usage?
• RQ3 How does the proposed method compare to segmentation and centerline detection in preserving junctions and continuity?
• RQ4 To what extent does the method generalize across diverse datasets with varying texture and structure complexity?
• RQ5 Can the method achieve high F1 scores while using significantly fewer pixels than state-of-the-art approaches?

Key findings

• The proposed method achieved the highest F1 score on the RecA (0.59), Aerial (0.59), and Cracks (0.38) datasets, outperforming all competing methods.
• On the DRIVE dataset, the method achieved an F1 score of 0.36, slightly below the best-performing method, due to fixed thickness parameter in binary pattern modeling.
• The proposed method used only 0.17% of pixels on the DRIVE dataset, significantly lower than Sironi et al. (5.41%) and Becker et al. (12.67%), indicating superior efficiency.
• The algorithm successfully reconstructed fine branches and junctions in retinal and crack images, as shown in intermediate reconstruction steps (Figure 6), demonstrating progressive topological detail recovery.
• Despite using a relatively small and imperfect training dataset, the method achieved robust performance across all datasets, including those with high texture complexity like Cracks.
• The method's reliance on structured ranking and geodesic path selection enables faithful preservation of topological features even under partial occlusion or low contrast.