Abstract
<title>Abstract</title> <p>Image classification remains a fundamental problem in computer vision, with convolutional neural networks (CNNs) traditionally providing strong performance because of their image-specific inductive biases. More recently, Vision Transformers (ViTs) have demonstrated that self-attention can also learn effective visual representations by processing images as sequences of patches. This study compares three deep-learning approaches for CIFAR-10 image classification: a Vision Transformer, a hybrid CNN–Multi-Layer Perceptron (CNN–MLP), and an ImageNet-pretrained ResNet-18 adapted through transfer learning. The models are evaluated using classification accuracy, precision, recall, F1-score, training and validation loss behavior, and a confusion-matrix analysis of the Vision Transformer. The ViT achieves 78.2% accuracy on the 500-image evaluation represented by its confusion matrix, with macro precision, recall, and F1-score of approximately 77.8%, 77.8%, and 77.7%, respectively. The hybrid CNN–MLP achieves 84.2% accuracy, whereas the transfer-learned ResNet-18 achieves the highest reported accuracy of 88.7%. The results indicate that the ViT can learn meaningful representations from low-resolution images, but convolutional inductive biases and large-scale pretraining provide a substantial advantage in the limited-data setting considered in this work. Vision Transformer, CIFAR-10, image classification, convolutional neural network, ResNet-18, transfer learning, self-attention, deep learning</p>