Year: 2025
Type: article
Abstract: This study aims to validate the performance differences between Convolutional Neural Networks and Transformer architectures in image classification. Based on their distinct feature extraction mechanisms, three algorithms—Resnet101, Resnet152, and Vision Transformer (ViT)—were trained and tested on the mini-Imagenet dataset. The results show that ViT achieved a top-1 accuracy of 92.79%, surpassing Resnet101's 90.98% and Resnet152's 91.71%. This largely demonstrates ViT's superiority over traditional convolutional neural networks in classification accuracy, as its unique Transformer architecture can more effectively capture global features and contextual information. Compared to the limitations of conventional CNN algorithms, ViT evidently exhibits enhanced performance in complex image classification tasks through its self-attention mechanism.
Author: Yiming Gou
Institution: Chengdu University of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4410350177