Image Clustering Algorithm Based on Predefined Evenly-Distributed Class Centroids and Composite Cosine Distance.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural network-based clustering algorithms aim to optimize feature representation for accurate image clustering.
• Existing methods struggle with low accuracy when dealing with complex natural images.

Purpose of the Study:

• To develop an improved image clustering algorithm for complex natural images.
• To enhance clustering accuracy by introducing predefined evenly-distributed class centroids (PEDCC) and composite cosine distance.

Main Methods:

• Designed an encoder-only network structure with normalized latent features.
• Replaced Euclidean distance with composite cosine distance in two latent feature space loss functions.
• Utilized contrastive learning and a self-supervised pretrained model.

Main Results:

• Contrastive learning significantly improves latent feature learning quality.
• Composite cosine distance is more suitable than Euclidean distance for normalized latent features and PEDCC-based Maximum Mean Discrepancy (MMD) loss.
• Self-supervised pretraining effectively boosts clustering performance on complex natural images.

Conclusions:

• The proposed image clustering algorithm achieves superior performance compared to state-of-the-art methods.
• The combination of normalized latent features, composite cosine distance, and PEDCC is effective for image clustering.
• Contrastive learning and self-supervised pretraining are crucial for enhancing clustering accuracy in deep learning models.