Comparative Study of Transformers for Image Processing

Transformers, neural networks based purely on self-attention mechanisms, have achieved great success in natural language processing. They have become the standard solution in this domain. BERT and GPT are two examples of state-of-the-art transformers. In addition, transformers have also attracted much attention in image processing in recent years. In the current research, there are several approaches, none of which has yet been able to assert itself. In this thesis we explain the attention mechanism and how it is used in the transformer architecture. Using selected examples, we show how the architecture is transferred to the image domain. We compare different approaches with each other and with the current state-of-the-art solution, convolutional neural networks. Furthermore, we visualize the attention within the transformer to gain insights into the decisions of the network.