Image Clustering via Deep Embedded Dimensionality Reduction and Probability-Based Triplet Loss.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Image clustering presents significant challenges compared to image classification, particularly due to the lack of supervised information for deep learning models.
• Existing methods struggle with high-dimensional image data, effective feature extraction, and the integrated process of feature extraction, dimensionality reduction, and clustering.

Purpose of the Study:

• To propose a novel framework, Deep Embedded Dimensionality Reduction Clustering (DERC), to address the core challenges in image clustering.
• To develop a unified framework that effectively combines image embedding, dimensionality reduction, and clustering.
• To enhance image clustering accuracy by integrating a novel probability-based triplet loss with reconstruction loss.

Main Methods:

• Introduced the Deep Embedded Dimensionality Reduction Clustering (DERC) framework.
• Developed a novel probability-based triplet loss function for retraining the DERC network.
• Integrated reconstruction loss and the probability-based triplet loss within a unified framework.

Main Results:

• The proposed DERC framework effectively addresses the curse of dimensionality, feature extraction, and integrated clustering challenges.
• The novel probability-based triplet loss significantly improves image clustering accuracy.
• Experimental results demonstrate that DERC outperforms state-of-the-art methods on multiple benchmark datasets.

Conclusions:

• DERC is the first framework to effectively combine image embedding, dimensionality reduction, and clustering.
• The integration of probability-based triplet loss and reconstruction loss offers a powerful approach to enhance image clustering.
• The proposed method shows superior performance, advancing the field of unsupervised image clustering.