Learning a Deep Fourier Attention Generative Adversarial Network for Light Field Image Super-Resolution

Human eyes can see the three-dimensional (3D) world because they receive the light emitted by objects, and the light field (LF) is a complete representation of the set of light in the 3D world. Light field image super resolution (LFISR) aims at reconstructing high-resolution LF images from their low-resolution counterparts which captured by the LF camera. In recent years, although convolutional neural networks (CNNs) can bring good performance to LFISR tasks, they cannot recover more real finer texture details. Therefore, it is a difficult challenge to generate realistic images of LF that can satisfy human perception. To address these problems, we propose a novel LFISR model by learning a deep Fourier attention generative adversarial network (GAN). Specifically, the generator network and loss function of the traditional single image super-resolution reconstruction (SISR) GAN model are improved for LFISR reconstruction. Furthermore, we design an attention module based on deep Fourier channel, which leverages the frequency content difference across distinct features to learn precise hierarchical representations of high-frequency information of diverse structures. Extensive experiments are performed on the mainstream LF datasets, leading to the state-of-the-art results of our method. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.