Progressive Growth-Based Momentum Contrast for Unsupervised Representative Learning in Classification Tasks

Contrastive unsupervised learning has made significant progress, but there is still potential for improvement by capturing finer details in input data. In this letter, we present PGMoCo, a Progressive Growth-based Momentum Contrast framework for unsupervised representative learning in classification tasks. PGMoCo begins by learning the overall distribution of samples at a coarse scale and progressively refines the representation by incorporating increasingly finer details. PGMoCo consists of data augmentation, progressive growth, an alternative multilayer perceptron (MLP) head, and a loss function. First, PGMoCo applies transformation-based data augmentation to the input samples. Then, it progressively learns features at multiple scales, uses an alternative MLP head to project latent representations into a contrastive loss space, and finally employs a specialized loss function to classify the samples. We evaluate PGMoCo on three datasets: CIFAR-10 and PolyU Palmprint (image classification) and H-MOG (person identification). PGMoCo achieves classification accuracies of 86.76% on CIFAR-10, 95.94% on PolyU Palmprint, and 80.10% on H-MOG, outperforming existing state-of-the-art methods.