From Convolution to Attention: A Historical Review and Comparative Analysis of Deep Learning Models in Image Classification

Image classification, a core computer vision task, profoundly impacts a computer system’s visual understanding. This paper comprehensively reviews deep - learning - driven image classification methods, focusing on the competition and collaborative evolution between CNNs and Transformer architectures. By comparing the performance of classic CNNs (e.g., ResNet, EfficientNet) and cutting - edge Transformer models (e.g., ViT, Swin Transformer) on benchmark datasets like ImageNet and CIFAR, it reveals how innovations like residual connections and attention mechanisms enhance classification accuracy, computational efficiency, and generalization. Experiments show hybrid architectures outperform same - scale CNNs, proving the advantage of combining local inductive bias with global dynamic modeling. Also, self - supervised learning shows potential in few shot scenarios, offering a solution for low - data situations. This paper further explores the role of neural architecture search and quantized training in model lightweight and robustness, supporting next - generation image classification technology.