Image Super-Resolution Using Light-Weight Deep Learning Methods
Honghui Zhu, Zhuqing Jiang, Aidong Men · 2020
Recently, design of neural networks with compact model size, light computation cost and high performance, has attracted much attention due to the need of applications on mobile devices. Inspired by knowledge distillation, we design a multi-level supervised compact network (MSCN) for SR with lightweight parameters and computational complexity. In the proposed MSCN, we employ a compact model which is not only supervised by the HR label, but also supervised by the intermediate representation of a cumbersome model. We comprehensively evaluate our MSCN on benchmark datasets with two widely used image quality metrics: PSNR and SSIM. The results show that our proposed method outperforms the compact network which is without multi-level supervisions by a comparable margin in terms of PSNR and SSIM.