Learning clustering-friendly representations via partial information discrimination and cross-level interaction.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning
• Clustering Algorithms

Background:

• Existing deep clustering methods often neglect sample-wise relationships and focus on full-image representations.
• Current approaches overlook discriminative information in partial image regions and lack cross-level interaction.
• Limitations in deep clustering hinder optimal representation learning and clustering accuracy.

Purpose of the Study:

• To introduce a novel deep image clustering approach, PICI (Partial Information and Cross-level Interaction).
• To overcome limitations of existing deep clustering methods by incorporating partial information and cross-level interactions.
• To enhance representation learning and clustering performance through a unified framework.

Main Methods:

• Utilized a Transformer encoder backbone with two augmentation types for parallel views.
• Integrated masked patches and class tokens for processing through the Transformer encoder.
• Employed three joint modules: Partial Information Self-Discrimination (PISD) for reconstruction, Partial Information Contrastive Discrimination (PICD) for contrastive learning, and Cross-Level Interaction (CLI) for consistency.

Main Results:

• The PICI approach demonstrated superior performance across six image datasets compared to state-of-the-art methods.
• Achieved an accuracy (ACC) of 0.772 on the RSOD dataset and 0.634 on the UC-Merced dataset.
• Showcased significant improvements of 29.7% and 24.8% over the best baseline on RSOD and UC-Merced, respectively.

Conclusions:

• The PICI approach effectively bridges masked image modeling and deep contrastive clustering.
• Offers a novel pathway for enhanced representation learning and improved clustering accuracy.
• The proposed method shows significant advancements in deep image clustering tasks.