Image Classification via an Improved Vision Transformer: Enhancing Global and Local Feature Modeling

Abstract

This study proposes a method for image classification based on an improved Vision Transformer to address the limitations of traditional convolutional neural networks in global modeling and long-range dependency capture. The input images are first divided into patches and mapped into embeddings to ensure the preservation of local details during serialization. In the encoder, multi-head self-attention and feed-forward networks are introduced to enhance cross-region feature interaction, while residual connections and normalization alleviate gradient vanishing in deep layers to achieve stable and efficient feature learning. At the classification stage, an aggregation function is used to combine patch representations into a global feature, followed by a fully connected layer for final prediction. The CIFAR-100 dataset, covering diverse fine-grained categories, is used to evaluate the adaptability of the model in complex scenarios. Systematic comparisons and sensitivity analyses show that the method outperforms others in AUC, ACC, Precision, and Recall, demonstrating clear advantages in feature modeling and classification performance. This research enriches the theoretical exploration of Vision Transformer optimization and provides a robust and efficient solution for image classification tasks.