Image Clustering Using an Augmented Generative Adversarial Network and Information Maximization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional clustering algorithms struggle with high-dimensional image features and unprocessed data.
• Performance degradation is common in image clustering due to feature dimensionality and distance function sensitivity.
• The rise of unlabeled datasets necessitates advanced clustering techniques.

Purpose of the Study:

• To propose a robust deep clustering framework for unlabeled image datasets.
• To enhance feature representation learning for improved clustering performance.
• To address the limitations of traditional clustering methods in high-dimensional image spaces.

Main Methods:

• A modified generative adversarial network (GAN) with Sobel operations in the discriminator to improve feature separability.
• An auxiliary classifier trained to maximize mutual information between discriminator-generated representations.
• Introduction of a penalty term for robustness and a tolerance hyper-parameter for imbalanced data.

Main Results:

• The proposed deep clustering framework achieves competitive performance against state-of-the-art methods.
• Effective clustering demonstrated on benchmark datasets like CIFAR-10, CIFAR-100/20, and STL10.
• Enhanced feature separability and robustness contribute to superior clustering outcomes.

Conclusions:

• The novel deep clustering framework effectively handles challenges posed by high-dimensional image features.
• The integration of modified GANs and auxiliary classifiers offers a promising direction for unsupervised image clustering.
• The method shows significant potential for applications involving large unlabeled image collections.