Image restoration refinement with Uformer GAN

Xu Ouyang, Ying Chen, Kaiyue Zhu, Gady Agam

Abstract: In this paper, we propose a novel approach for image restoration refinement, aiming to refine the result of restoring a clear original image from a noisy or blurry one. Our proposed method, Uformer GAN, combines the use of Transformer blocks and restoration refinement to achieve superior performance in image restoration tasks. The generator in our Uformer GAN model comprises Transformer blocks followed by a convolution layer. This design allows the model to learn the connections among each pixel of an image and capture local context features. The discriminator, on the other hand, consists of Transformer blocks and convolution blocks to balance the model’s capability and efficiency. Additionally, instead of adopting multi-stage networks like other image restoration methods and training them concurrently, we solely focus on training a post-processing network for refined image restoration. This approach reduces the complexity of the overall image restoration process and ensures that our refinement is scalable to various image restoration techniques. We demonstrate the effectiveness of our proposed methods on two datasets: the image deblurring GOPRO dataset and the image denoising SIDD dataset. Our approach shows superior performance compared to other state-of-the-art methods in both datasets.