Abstract
This paper presents a comprehensive comparative study on transfer learning and Vision Transformers for Kazakhstan banknote recognition, aimed at assisting visually impaired individuals. We systematically evaluate four state-of-the-art CNN architectures (ResNet50, VGG16, MobileNetV2, GoogleNet) pretrained on ImageNet alongside Vision Transformer (ViT-B/16), training on a dataset of 1,259 images of Kazakhstani banknotes captured under diverse real-world conditions. Through rigorous hyperparameter optimization across three random seeds with proper train-validation-test separation, we find that Vision Transformers achieve exceptional performance$(99.21 \pm 0.79 \%$test accuracy) with perfect 100% validation accuracy using RMSprop lr=0.001 and Adam/RMSprop lr=0.01, followed by ResNet50$(96.06 \pm 0.79 \%)$which achieved 99.73% validation accuracy with SGD lr=0.01. MobileNetV2 achieved$\mathbf{9 5. 0 1} \pm \mathbf{0. 4 5} \boldsymbol{\%}$accuracy, while GoogleNet and VGG16 obtained$92.39 \pm 1.64 \%$and$91.86 \pm 2.27 \%$respectively. Our findings advance banknote recognition technology while providing insights into the comparative effectiveness of CNNs versus transformers on limited-data computer vision tasks.