Boundary-complexity-guided adaptive weighting for single-image unsupervised superpixel segmentation
Abstract
Superpixel segmentation provides an important mid-level representation for computer vision, but accurate boundary adherence remains difficult in images with weak contours, cluttered backgrounds, and complex textures. We propose a boundary-complexity-guided adaptive weighting framework for single-image unsupervised superpixel segmentation, termed AW-Super. The proposed Enhanced Boundary Feature Extraction Module (EBFEM) fuses semantic edge cues with gradient-based local structural information to produce a continuous complexity map, which is then used to reweight the optimization objective spatially. As a result, the method imposes stronger boundary-sensitive constraints in structurally complex regions while encouraging shape regularity in relatively homogeneous areas. We further introduce an adaptive stopping criterion into the Deep Image Prior optimization process, replacing a fixed iteration schedule with an image-dependent termination rule. Experiments on BSDS500 and SBD under matched superpixel budgets ( N=50 –500) demonstrate high boundary recall together with competitive overall segmentation performance, without requiring dataset-level training or ground-truth superpixel annotations. An additional cross-domain evaluation on DRIVE retinal images against SLIC and SEEDS further supports the applicability of AW-Super beyond natural-image benchmarks.
Related
- Contrastive adversarial mix-up learning for unsupervised cross-modality image segmentation
- CalciSeg : A versatile approach for unsupervised segmentation of calcium imaging data
- MIRTO: a registration-gated, multiverse-tested evaluation protocol for unsupervised anomaly segmentation in brain MRI
- A Conditional, Boundary-Aware, and Auditable Deletion-Effect Weighting Framework for MCDM with Positive Decision Matrices and a Fixed Set of Alternatives
- Evaluation of boundary-aware UNet++ for low-grade glioma segmentation
BibTeX
@article{wang2026boundarycomplexityguided,
title = {{Boundary-complexity-guided adaptive weighting for single-image unsupervised superpixel segmentation}},
author = {Nannan Wang and Qianyi Zhang and Yongxia Zhang and Wei Wei},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-74191-2},
url = {https://www.nature.com/articles/s41598-026-74191-2},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Boundary-complexity-guided adaptive weighting for single-image unsupervised superpixel segmentation
AU - Nannan Wang
AU - Qianyi Zhang
AU - Yongxia Zhang
AU - Wei Wei
PY - 2026
DA - 2026/10/06
JO - Scientific Reports
DO - 10.1038/s41598-026-74191-2
UR - https://www.nature.com/articles/s41598-026-74191-2
AB - Superpixel segmentation provides an important mid-level representation for computer vision, but accurate boundary adherence remains difficult in images with weak contours, cluttered backgrounds, and complex textures. We propose a boundary-complexity-guided adaptive weighting framework for single-image unsupervised superpixel segmentation, termed AW-Super. The proposed Enhanced Boundary Feature Extraction Module (EBFEM) fuses semantic edge cues with gradient-based local structural information to produce a continuous complexity map, which is then used to reweight the optimization objective spatially. As a result, the method imposes stronger boundary-sensitive constraints in structurally complex regions while encouraging shape regularity in relatively homogeneous areas. We further introduce an adaptive stopping criterion into the Deep Image Prior optimization process, replacing a fixed iteration schedule with an image-dependent termination rule. Experiments on BSDS500 and SBD under matched superpixel budgets ( N=50 –500) demonstrate high boundary recall together with competitive overall segmentation performance, without requiring dataset-level training or ground-truth superpixel annotations. An additional cross-domain evaluation on DRIVE retinal images against SLIC and SEEDS further supports the applicability of AW-Super beyond natural-image benchmarks.
SN - 2045-2322
ER -