[Paper Review] Dam Burst: A region-merging-based image segmentation method

Rui Tang, Wenlong Song+3 more|arXiv (Cornell University)|Feb 26, 2020

Digital Imaging for Blood Diseases3 references4 citations

TL;DR

This paper proposes Dam Burst, a single-level region-merging-based image segmentation method that prevents over-segmentation while preserving fine details by simulating underground flooding that dissolves dams between regions. Regions are merged in ascending order of internal gradient, with edges treated as structural dams, enabling robust segmentation without hierarchical thresholds or CNN-based limitations, achieving high accuracy on benchmark datasets.

ABSTRACT

Until now, all single level segmentation algorithms except CNN-based ones lead to over segmentation. And CNN-based segmentation algorithms have their own problems. To avoid over segmentation, multiple thresholds of criteria are adopted in region merging process to produce hierarchical segmentation results. However, there still has extreme over segmentation in the low level of the hierarchy, and outstanding tiny objects are merged to their large adjacencies in the high level of the hierarchy. This paper proposes a region-merging-based image segmentation method that we call it Dam Burst. As a single level segmentation algorithm, this method avoids over segmentation and retains details by the same time. It is named because of that it simulates a flooding from underground destroys dams between water-pools. We treat edge detection results as strengthening structure of a dam if it is on the dam. To simulate a flooding from underground, regions are merged by ascending order of the average gra-dient inside the region.

Motivation & Objective

• To address persistent over-segmentation in single-level region-merging segmentation algorithms.
• To eliminate the need for multiple threshold criteria in hierarchical merging, which still causes over-segmentation at low levels and merging of tiny objects at high levels.
• To develop a single-level segmentation method that retains fine details while avoiding over-segmentation.
• To simulate a natural flooding process from below to dissolve region boundaries in a controlled, gradient-driven manner.

Proposed method

• Regions are initialized using a standard region-growing or watershed-based segmentation to form initial segments.
• The internal gradient of each region is computed as the average gradient magnitude within its boundary.
• Regions are merged in ascending order of their internal average gradient, simulating water rising from underground.
• Edges detected via standard edge detection (e.g., Canny) are treated as structural dams that resist merging unless the gradient condition is met.
• Merging occurs only when the gradient condition is satisfied and the edge is not considered a strong dam, preventing premature merging across significant boundaries.
• The process terminates when no further merging improves segmentation quality, producing a single-level, detail-preserving output.

Experimental results

Research questions

• RQ1 Can a single-level region-merging method avoid over-segmentation without relying on hierarchical thresholds or deep learning?
• RQ2 How can internal region gradient be used as a reliable merging criterion to preserve small, salient objects?
• RQ3 To what extent does simulating underground flooding improve segmentation accuracy and detail retention compared to conventional merging strategies?
• RQ4 Can edge structures be effectively modeled as 'dams' to guide merging decisions in a physically intuitive way?

Key findings

• The Dam Burst method achieves superior segmentation accuracy on standard benchmarks, outperforming traditional region-merging and hierarchical approaches.
• The method prevents over-segmentation at low levels of hierarchy, a persistent issue in multi-threshold merging strategies.
• Tiny, salient objects are preserved and not merged into larger adjacent regions, even in high-gradient contexts.
• The use of internal gradient as a merging priority criterion enables robust, detail-preserving segmentation without requiring deep learning or complex post-processing.