|
Ultra-high-definition (UHD) image super-resolution (SR) has attracted increasing attention due to the popularity of modern devices, such as smartphones, which support the capture of UHD images, e.g. 4K and 8K images. However, existing UHD SR methods process the image in the spatial domain only. This limitation hinders their ability to effectively utilize the rich details and fine-grained textures in local areas of UHD images. To address this issue, our proposed method comprehensively exploits the global and local features of UHD images by combining spatial and frequency features. Additionally, previous UHD image SR methods can only handle a fixed scaling factor, but real-world applications very often require upscaling low-resolution images with different scales. Therefore, we employ an arbitrary-scale strategy in the SR training process, enabling super-resolution of UHD images at any scale with a single trained model. Experimental results demonstrate the effectiveness and superiority of our proposed method.
|
