Review of Modern Image Classification: Paradigms, Architectures, and the Post-Supervised Era

A comprehensive literature review on Image Classification, contrasting the resurgence of ConvNeXt architectures against Vision Transformers (ViTs) alongside the paradigm shift toward Self-Supervised Learning.

Convolutional Neural Networks or CNNs, have been firmly established as state-of-the-art approaches in image classification. Since then, there have been numerous attempts to push the boundaries of image classification and deep learning algorithms. The rise of Self-Supervised Learning (SSL), characterized by contrastive methods and Masked Image Modeling (MIM), alongside language-supervised foundation models like CLIP (Contrastive Language-Image Pre-training), has fundamentally altered the objective functions of computer vision. This review combines modern advances in image classification, assessing the progression from classic supervised techniques to deep learning-driven models and emerging post-supervised approaches. Key architectural innovations such as CNNs, transformers, and hybrid systems are critically examined alongside changes in data curation and evaluation strategy. Special attention is given to recent trends enabling robust performance with limited supervision, including self-supervised and few-shot learning. Finally, we analyze the scaling laws that now govern the performance of these massive systems, elucidating the complex trade-offs between accuracy, robustness, and efficiency in a world where the boundaries between vision, language, and diverse modalities are increasingly dissolved.