Comparative Study of Vision Transformer (ViT), Swin Transformer, and Transformer in Transformer (TNT) on Image Classification

The advent of transformer models in computer vision has revolutionized image classification, significantly improving performance compared to standard convolutional neural networks (CNNs). Among these models, the Vision Transformer (ViT), Swin Transformer, and Transformer in Transformer (TNT) represent three leading architectures that apply self-attention mechanisms to image data. In this paper, we perform a comparative evaluation of these models using the CIFAR-10 dataset, focusing on classification performance, computational complexity, and training efficiency. We assess each model using accuracy, precision, recall, F1-score, AUC, and confusion matrices. ViT provides a simple yet powerful mechanism for leveraging global context in image understanding. Swin Transformer introduces a hierarchical design and shifted windows to improve computational efficiency. TNT enhances the process further by enabling fine-grained local and global attention through inner- patch and patch tokenization. Our findings indicate that while TNT achieves the highest overall accuracy and robustness, Swin Transformer maintains an excellent balance between accuracy and computational efficiency on smaller datasets like CIFAR-10.