Memory-Efficient Discrete Cosine Transform Domain Weight Modulation Transformer for Arbitrary-Scale Super-Resolution

Abstract

Recently, several arbitrary-scale models have been proposed for single-image super-resolution. Furthermore, the importance of arbitrary-scale single image super-resolution is emphasized for applications such as satellite image processing, HR display, and video-based surveillance. However, the baseline integer-scale model must be retrained to fit the existing network, and the learning speed is slow. This paper proposes a network to solve these problems, processing super-resolution by restoring the high-frequency information lost in the remaining arbitrary-scale while maintaining the baseline integer scale. The proposed network extends an integer-scaled image to an arbitrary-scale target in the discrete cosine transform spectral domain. We also modulate the high-frequency restoration weights of the depthwise multi-head attention to use memory efficiently. Finally, we demonstrate the performance through experiments with existing state-of-the-art models and their flexibility through integration with existing integer-scale models in terms of peak signal-to-noise ratio (PSNR) and similarity index measure (SSIM) scores. This means that the proposed network restores high-resolution (HR) images appropriately by improving the image sharpness of low-resolution (LR) images.

Keywords

• machine learning
• pattern recognition
• arbitrary scale super-resolution
• DCT spectral domain
• multi-head self attention
• vision transformer