.png){kind=icon}
.png){kind=icon}
Return
DeepCut plus plus : Graph-based unsupervised segmentation with feature fusion and diffusion learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.114975.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Unsupervised image segmentation remains a fundamental yet challenging task due to the absence of ground-truth labels and the complexity of semantic structures in visual data. Despite significant advancements, recent graph-based unsupervised segmentation methods face important limitations. These models exchange information only among neighboring nodes, which restricts them to capturing local structures while overlooking long-range relationships between distant image regions. The absence of a global feature-smoothing mechanism often leads to fragmented or noisy segments. Moreover, without integrating features across multiple scales, such models struggle to accurately delineate object boundaries. To address these challenges, we present DeepCut++, which introduces three key contributions: (1) a Personalized PageRank (PPR) diffusion module that acts as a learnable low-pass filter, enabling global feature propagation while suppressing high-frequency noise; (2) a multi-scale feature fusion mechanism that combines fine-grained spatial details with high-level semantic features; and (3) an end-to-end trainable framework unifying these components with clustering objectives. Achieving CorLoc of 70.59% and 72.74% on VOC 2007/2012 for object localization; mIoU of 79.49%, 61.49%, and 76.84% on CUB, DUTS, and ECSSD for single object segmentation; and 45.12 NMI / 21.9 ARI on CUB for semantic part segmentation clearly demonstrates the consistent superiority of DeepCut++ over 24 state-of-the-art methods.
Keywords:
Unsupervised segmentation
Personalized PageRank
Feature fusion