Imbalcbl: addressing deep learning challenges with small and imbalanced datasets

Abstract

Deep learning, while transformative for computer vision, frequently falters when confronted with small and imbalanced datasets. Despite substantial progress in this domain, prevailing models often underachieve under these constraints. Addressing this, we introduce an innovative contrast-based learning strategy for small and imbalanced data that significantly bolsters the proficiency of deep learning architectures on these challenging datasets. By ingeniously concatenating training images, the effective training dataset expands from n to \(n^2\), affording richer data for model training, even when n is very small. Remarkably, our solution remains indifferent to specific loss functions or network architectures, endorsing its adaptability for diverse classification scenarios. Rigorously benchmarked against four benchmark datasets, our approach was juxtaposed with state-of-the-art oversampling paradigms. The empirical evidence underscores our method’s superior efficacy, outshining contemporaries across metrics like Balanced accuracy, F1 score, and Geometric mean. Noteworthy increments include 7–16% on the Covid-19 dataset, 4–20% for Honey bees, 1–6% on CIFAR-10, and 1–9% on FashionMNIST. In essence, our proposed method offers a potent remedy for the perennial issues stemming from scanty and skewed data in deep learning.