Image Clustering: Utilizing Teacher-Student Model and Autoencoder

Complete and high-quality labeled dataset is indispensable for image classification. Considering the often arduous task of data labeling, clustering algorithms are commonly utilized in the preliminary stages of data classification to perform preliminary categorization. Traditional clustering algorithms, such as K-means and Gaussian Mixture Model, often struggle to effectively cluster images. In this paper, we propose a novel image clustering method utilizing teacher-student model and autoencoder. Specifically, the teacher model is a fully connected autoencoder and the student model consists of multiple convolutional autoencoders. We firstly obtain sub-datasets by applying the K-means algorithm to cluster the output of the teacher model, then utilize the student model to achieve clustering through an iterative data exchange method where the same data can converge. In addition, the proposed method can filter low-quality data to a certain extent by recording the frequency of exchanges because clustering is achieved during multiple exchanges of data. Experiments on a synthetic dataset and three benchmark image datasets (MNIST, FashionMNIST and USPS) show that our method can achieve satisfactory clustering results.