Dynamic image clustering from projected coordinates of deep similarity learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional clustering algorithms often require pre-defined parameters like the number of clusters.
• Density-based clustering methods are unsuitable for high-dimensional datasets.
• Recent advancements integrate deep similarity learning with cluster assignment.

Purpose of the Study:

• To propose a novel framework for dynamic image clustering without prior knowledge of the cluster count.
• To enable effective clustering of high-dimensional data, specifically images.
• To facilitate consistent prediction of unseen data without retraining.

Main Methods:

• A deep learning model is employed to learn data similarity from scratch.
• A coordinate learning model projects high-dimensional data into a two-dimensional space.
• A novel raster clustering algorithm is introduced for evaluating and classifying projected data.

Main Results:

• The proposed framework successfully performs dynamic image clustering without specifying the number of clusters.
• High-dimensional data, such as images, can be effectively clustered.
• The mechanism allows for consistent prediction of unseen data without re-training or consolidation with training data.

Conclusions:

• The developed framework offers a robust solution for unsupervised clustering of high-dimensional image data.
• It overcomes limitations of existing methods by eliminating the need for parameter pre-definition and handling high dimensionality.
• The approach demonstrates potential for real-world applications requiring dynamic and adaptive data analysis.