Image Super-Resolution Using GAN Based Vision Transformer

The goal of image super-resolution is to create a high-resolution image from a low-resolution one. Deep learning has significantly advanced this task, enabling the generation of high-quality images. However, the increasing complexity of neural networks leads to greater computational and memory demands, posing challenges for efficiency and real-time applications. It remains an active field with the potential to overcome many applications' resolution constraints. Transformers' strong self-attention mechanism has allowed it to achieve notable advancements in computer vision challenges in recent years. Due to the high computational demands of transformers, there are currently no widely adopted vision transformer networks designed explicitly for highly high-resolution images. This remains a challenge in current research. We suggested a comprehensive vision transformer-based GAN network to solve this issue. This network has two types of transformer networks: transformer discriminator and transformer generator. In order to produce 2x and 4x resolution images using image patches as input, we present the innovative Transformer Encoder Generator module. It can gradually reduce GPU memory usage and achieve excellent resolution.