Transformer and Multi-Scale Feature Extraction for Image Classification

College of Intelligence and Computing, Tianjin University, Tianjin, China

yangyang_420@tju.edu.cn

Abstract

Image classification refers to enabling computers to automatically identify and classify the categories of input images. With the continuous optimization of the Convolutional Neural Network (CNN)structure and its in-depth application in image classification tasks, the classification accuracy and computational efficiency of in this field have been significantly improved. However, CNN still has limitations in its ability to handle long-distance dependencies in images, and its ability to capture elements with large distances is still lacking, resulting in decreased performance in image classification. In response to this problem, this paper adds a multi-scale feature extraction mechanism based on the Vision Transformer (ViT). The experimental results show that the model with a multi-scale feature extraction mechanism has a substantial improvement in performance, with the accuracy increased from 89.42% to 98.39%. This method can not only effectively enhance long-range dependency modeling and shorten the connection path for long-range semantic relationships, but also improve the information integration efficiency of the Transformer, increase the recognition accuracy of targets, and simultaneously enhance sensitivity to edge and local information.