Vision Transformers Versus Convolutional Neural Networks: Comparing Robustness by Exploiting Varying Local Features

Robustness of deep learning models is a crucial point that should be evaluated in computer vision tasks. Deep learning models must be able to produce the same results under different conditions such as brightness, image rotation, etc. Deep learning models must have high level robustness to be reliable to address real world problems. Unreliable systems may be harmful for critical systems such as security, healthcare, and military. In this study, it is aimed to evaluate image rotation robustness of Convolutional Neural Networks (CNNs) and Transformer-based deep learning models: MobileNetV2, Residual Network 18 (ResNet18), Visual Geometry Group 16 (VGG16), EfficientNetB1, Vision Transformer (ViT), Data-Efficient Image Transformers (DeiT), and Pooling-based Vision Transformer (PiT). Canadian Institute for Advanced Research 100 classes (CIFAR100), Canadian Institute for Advanced Research 10 classes (CIFAR10) and California Institute of Technology 256 classes (Caltech256) datasets were used to evaluate the performance of the models. All deep learning models were trained with original train datasets. Image embeddings of train, test datasets, and their rotated forms were extracted. A Support Vector Classifier (SVC) was trained with embeddings of original dataset and concatenation of original and rotated form of train dataset. Performances on original test dataset and their rotated forms were evaluated. 10, 50, and 90 degrees rotated form of the datasets were used for rotated test dataset. F1-score metrics were used compare model performances. In conclusion, Transformer-based classification algorithms are more robust to image rotation than the CNN-based models.