Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Data augmentation methods for Vision Transformers

Georgakilas Christos

Full record

Year 2022
Type of Item Diploma Work
Bibliographic Citation Christos Georgakilas, "Data augmentation methods for Vision Transformers", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2022
Appears in Collections


The Transformer architecture was first introduced in 2017 and has since become the standard for Natural Language Processing tasks, replacing Recurrent Neural Networks. For the first time, in 2021, the Transformer architecture was used with great success for computer vision tasks, proving that a Vision Transformer can, under certain conditions, outperform Convolutional Neural Networks and become the state-of-the-art in image recognition. One of the main challenges being tackled by subsequent work on Vision Transformers is the need of the architecture for humongous amounts of data during pre-training in order to achieve state-of-the-art accuracy on the downstream task. Some works have addressed this by altering or adding parts to the original Vision Transformer architecture while others are using Self-Supervised Learning techniques to take advantage of unlabeled data. This thesis explores data augmentation methods for Vision Transformers with the goal to increase the model’s accuracy and robustness on classification tasks, with limited amounts of data. Our augmentation methods are based on the architecture’s characteristics such as the self-attention mechanism and the input of discrete tokens. All methods are tested for the benchmark classification datasets CIFAR-10 and CIFAR-100 using Supervised Learning and yield great results. When training with the same model hyperparameters, our best augmentation method improves the baseline’s accuracy on CIFAR-10 and CIFAR-100 by 1.98 % and 2.71 % respectively.

Available Files