Evaluating Pre-Trained CNNs and Vision Transformers in Image Classification

Abstract

This study compares pre-trained Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for emergency vehicle classification using a curated dataset of 2312 images. Pre-processing involved standardizing image dimensions and parsing XML annotations. Google Colab's GPUs were used to train models with the Adam optimizer and sparse categorical cross-entropy loss. The results revealed that ViTs are competitive with CNNs in classification performance. Among CNNs, MobileNetV1 achieved the highest accuracy of 92.21% with a training time of 32.47 seconds. ViTs, including ViT_B_32 and ViT_B_16, reached accuracies of 95.83% and 93.75%, respectively, but required longer training times. These findings, demonstrate the strength of ViTs in handling long-range dependencies while remaining competitive with traditional CNNs. This research provides valuable insights into architecture selection for image classification, contributing to the advancement of deep learning in emergency vehicle systems.