Leveraging variant of CAE with sparse convolutional embedding and two-stage application-driven data augmentation for image clustering
摘要
Existing deep clustering approaches often struggle with redundant feature learning, which limits their effectiveness. The primary goal of this study is to address these issues by developing a more robust deep clustering method. To achieve this, we propose a variant of the convolutional autoencoder (CAE) called SCDAC, which incorporates sparse convolutional embedding and a two-stage application-driven data augmentation approach. The proposed model operates in two main stages: pretraining and finetuning. In the pretraining stage, we employ application-driven data augmentation to train the CAE variant, focusing on learning robust features and constructing a foundational feature space using sparse convolutional embedding. During the finetuning stage, the model performs joint feature learning and cluster assignment. The feature learning task utilizes an augmented framework to control the input of both original and augmented data, preserving the local structure of images in the feature space. For cluster assignment, the framework controls the input of original data and uses the sparse convolutional embedding layer to obtain low-dimensional representations for soft cluster assignment. Experimental evaluations on six publicly available datasets demonstrate the effectiveness of the proposed model, with significant improvements in accuracy, particularly increases of