ARM: Any-Time Super-Resolution Method

This paper proposes an Any-time super-Resolution Method (ARM) to tackle the over-parameterized single image super-resolution (SISR) models. Our ARM is motivated by three observations: (1) The performance of different patches varies with SISR networks sizes. (2) There a tradeoff between computation overhead and reconstructed image. (3) Given input image, its edge information can be effective option estimate PSNR. Subsequently, we train supernet containing subnets sizes deal various complexity. To that effect, construct Edge-to-PSNR lookup table maps score patch PSNR for each subnet, together set costs subnets. In inference, are individually distributed better computation-performance tradeoff. Moreover, subnet shares weights supernet, thus no extra parameters introduced. setting multiple well adapt computational cost model dynamically available hardware resources, allowing task in service at any time. Extensive experiments on resolution datasets popular as backbones verify effectiveness versatility our ARM. source code https://github.com/chenbong/ARM-Net .