Image Clustering using an Augmented Generative Adversarial Network and Information Maximization

TL;DR

A modified generative adversarial network (GAN) and an auxiliary classifier consisting of a modified Sobel operations prior to the discriminator of the GAN to enhance the separability of the learned features and achieves competitive results compared with state-of-the-art clustering methods on a wide range of benchmark datasets.

Abstract

Image clustering has recently attracted significant attention due to the increased availability of unlabelled datasets. The efficiency of traditional clustering algorithms heavily depends on the distance functions used and the dimensionality of the features. Therefore, performance degradation is often observed when tackling either unprocessed images or high-dimensional features extracted from processed images. To deal with these challenges, we propose a deep clustering framework consisting of a modified generative adversarial network (GAN) and an auxiliary classifier. The modification employs Sobel operations prior to the discriminator of the GAN to enhance the separability of the learned features. The discriminator is then leveraged to generate representations as the input to an auxiliary classifier. An adaptive objective function is utilised to train the auxiliary classifier for clustering the representations, aiming to increase the robustness by minimizing the divergence of multiple representations generated by the discriminator. The auxiliary classifier is implemented with a group of multiple cluster-heads, where a tolerance hyper-parameter is used to tackle imbalanced data. Our results indicate that the proposed method significantly outperforms state-of-the-art clustering methods on CIFAR-10 and CIFAR-100, and is competitive on the STL10 and MNIST datasets.

Figures (7)

• Figure 1: Diagram of the proposal framework for clustering multi-dimensional datasets, where $G$ and $D$ stands for generator and discriminator model, respectively. $M$ denotes the discriminator's output which is directed to auxiliary classifier. $C$ represents the auxiliary classifier and $Y_{i}$ the multi-cluster heads.
• Figure 2: An illustration of the modified discriminator of the GAN framework in which two convolutional layers operate in front of the discriminator input. Initially, the three image layers (RGB) are converted into a gray-scale version via the application of pointwise layer. Afterwards a depthwise convolutional with a constant kernel implements the two directions of the Sobel's operations. The generated domain is concatenated with the original input and forwarded to the discriminator model.
• Figure 3: Clustering results on CIFAR-10 and MNIST. Visualization of the predictions made by the proposed clustering method, which includes the relevant images and their class probabilities (P(Y)).
• Figure 4: (a) and (b) present the accuracy of the three variants of the loss function based on a single cluster head model and multi-cluster heads, respectively. (c) illustrates the change of performance over different values of two scale factors $a_r$ and $a_{adv}$. For convenience $a_r$ is coloured and marked with different shapes, where the $x$ axis indicates the taken values of $a_{adv}$.
• Figure 5: Comparison of the discriminator's performance with or without the Sobel operation. The plots in the top row show the absolute accuracy achieved with the features extracted by k-means calculated in every 20 epochs (the blue line indicates the Sobel filter operation). The bottom row visualizes the errors of the discriminator and the generator, implying that adding the Sobel filters in the discriminator does not impair the training performance.