Scale-Aware Distillation Network for Lightweight Image Super-Resolution

Many lightweight models have achieved great progress in single image super-resolution. However, their parameters are still too many to be applied in practical applications, and it still has space for parameter reduction. Meanwhile, multi-scale features are usually underutilized by researchers, which are better for multi-scale regions’ reconstruction. With the renaissance of deep learning, convolution neural network based methods has prompted many computer vision tasks (e.g., video object segmentation [21, 38, 40], human parsing [39], human-object interaction detection [39]) to achieve significant progresses. To solve this limitation, in this paper, we propose a lightweight super-resolution network named scale-aware distillation network (SDNet). SDNet is built on many stacked scale-aware distillation blocks (SDB), which contain a scale-aware distillation unit (SDU) and a context enhancement (CE) layer. Specifically, SDU enriches the hierarchical features at a granular level via grouped convolution. Meanwhile, the CE layer further enhances the multi-scale feature representation from SDU by context learning to extract more discriminative information. Extensive experiments are performed on commonly-used super-resolution datasets, and our method achieves promising results against other state-of-the-art methods with fewer parameters.