ECCT: Efficient Contrastive Clustering via Pseudo-Siamese Vision Transformer and Multi-view Augmentation

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Image clustering is crucial for organizing unlabeled image data.
• Contrastive learning methods excel at learning discriminative features but struggle with global information and semantic continuity.
• Existing methods often have limited feature distributions, hindering contrastive learning's potential in clustering.

Purpose of the Study:

• To propose a novel deep clustering framework, Efficient Contrastive Clustering via Pseudo-Siamese Vision Transformer and Multi-view Augmentation (ECCT).
• To address the limitations of existing methods in capturing global information, preserving semantic continuity, and broadening feature distributions.
• To enhance the performance of image clustering using advanced deep learning techniques.

Main Methods:

• Introduced a pseudo-Siamese Vision Transformer (ViT) framework incorporating a Hilbert Patch Embedding (HPE) module for global feature extraction.
• Fused features from two ViT branches to achieve both global view and semantic coherence.
• Employed multi-view random aggressive augmentation to diversify feature distributions and learn richer contrastive features.

Main Results:

• ECCT demonstrated superior performance across five benchmark datasets compared to existing clustering methods.
• Achieved an Adjusted Rand Index (ARI) of 0.852 on the STL-10 dataset, outperforming the best baseline by 10.3%.
• On the ImageNet-Dogs dataset, ECCT reached an ARI of 0.424, an improvement of 4.8% over the best baseline.

Conclusions:

• ECCT effectively addresses the challenges of global information capture and semantic continuity in image clustering.
• The proposed framework significantly enhances the learning of comprehensive and rich contrastive features.
• ECCT represents a substantial advancement in deep clustering, offering improved accuracy and robustness.