adopts a trellis encoder-decoder network architecture, it can capture multi-scale features by integrating multiple decoding paths and exploits dense skip connections to obtain the supervised information. Yet, its shallow network structure cannot extract higher-level semantic information, which limits the ...
dently superiority of BSRDM demonstrates the importance of the noise and kernel modeling in SISR, since they both use the same network architecture to generate HR image even though BSRDM has fewer parameters (see Sec. 5.4). Two visual results on Gaussian noise ...