Enhancing Data Classification Across Multiple Domains: A Novel Two-Stage Self-Supervised Learning Approach
Kriti Bansal, Vikas Mittal · 2024
Self-supervised contrastive learning has emerged as a promising technique in computer vision. Self-supervised learning has become increasingly popular due to its capacity to sidestep the expense associated with annotating extensive datasets. This approach harnesses self-generated pseudo-labels for guidance and leverages the acquired representations across various subsequent tasks. The primary objective of contrastive learning is to ensure that representations of augmented versions of identical samples are clustered closely together while concurrently pushing apart embeddings from distinct samples. In this paper, we propose a novel two-stage methodology, SimCLR+XGB, for the detection of liver cancer and SAR classification. Despite having a limited number of labeled data, our approach achieves high accuracy and surpasses existing state-of-the-art models such as SimCLR+LR and BYOL. In our study, we introduce a self-supervised learning approach for data classification and work on multiple domains such as medical imaging and remote sensing, leveraging a large unlabeled dataset to learn representations of data. These learned representations are then utilized in the downstream tasks of liver cancer detection and SAR classification, resulting in efficient classification and reducing the burden of manual labeling. Our evaluation is conducted on two datasets: the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset and the Liver Tumor Segmentation (LiTS) dataset. Experimental results demonstrate significant improvements over previous methods, highlighting the effectiveness of our framework.