Size-Informed Representations for Unsupervised Image Clustering

Philipp Rajah Moura Srivastava, Charilaos Apostolidis, Saumya Pailwan, Patrick Koller, Leandros Stefanopoulos
September, 2025
Cite DOI IEEE Xplore
Proposed framework for size-informed deep image clustering using late-fusion size injection.

Abstract

Unsupervised image clustering aims to group unlabeled images into semantically meaningful categories without human supervision. While deep learning-based clustering frameworks such as semantic clustering by adopting neighbors (SCAN) and deep embedding clustering (DEC) have recently achieved promising results, the impact of preprocessing strategies—such as aspect-preserving padding and direct resizing—on clustering performance remains underexplored. This paper investigates both implicit size information (through preprocessing) and explicit size information injection in deep clustering frameworks, using a high-stakes, real-world dataset with significant image heterogeneity. We systematically evaluate SCAN and DINOv2-DEC clustering frameworks across multiple strategies: direct resizing, aspect ratio-preserving padding and late-fusion size injection with spectral encoding.

Type
Conference
Publication
2025 IEEE International Conference on Image Processing Workshops (ICIPW)

Philipp Rajah Moura Srivastava, Charilaos Apostolidis, Saumya Pailwan, Patrick Koller, and Leandros Stefanopoulos

Accepted and presented at the IEEE ICIP 2025 Satellite Workshop: “Generative AI for World Simulations and Communications & Celebrating 40 Years of Excellence in Education: Honoring Prof. Aggelos Katsaggelos,” Anchorage, Alaska, USA, Sept 14, 2025.

BibTeX Citation

@INPROCEEDINGS{Srivastava_2025_SizeInformedClustering,
  author={Srivastava, Philipp Rajah Moura and Apostolidis, Charilaos and Pailwan, Saumya and Koller, Patrick and Stefanopoulos, Leandros},
  booktitle={2025 IEEE International Conference on Image Processing Workshops (ICIPW)}, 
  title={Size-Informed Representations for Unsupervised Image Clustering}, 
  year={2025},
  volume={},
  number={},
  pages={743-748},
  keywords={Visualization;Image coding;Conferences;Semantics;Pipelines;Data models;Image preprocessing;Standards;Unsupervised Image Clustering;SelfSupervised Learning;Size Encoding;Image Preprocessing;Deep Embedded Clustering},
  doi={10.1109/ICIPW68931.2025.11386144}}
Unsupervised Image Clustering Self-Supervised Learning Size Encoding Image Preprocessing Deep Embedded Clustering SCAN DINOv2 IEEE ICIP
Patrick Koller
Patrick Koller
PhD Candidate

I work on scientific machine learning, focusing on physics-informed neural networks, neural operators, and large-scale physical simulations.