eduzhai > Applied Sciences > Engineering >

Frequency Domain-based Perceptual Loss for Super Resolution

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 6 pages

Abstract: We introduce Frequency Domain Perceptual Loss (FDPL), a loss function forsingle image super resolution (SR). Unlike previous loss functions used totrain SR models, which are all calculated in the pixel (spatial) domain, FDPLis computed in the frequency domain. By working in the frequency domain we canencourage a given model to learn a mapping that prioritizes those frequenciesmost related to human perception. While the goal of FDPL is not to maximize thePeak Signal to Noise Ratio (PSNR), we found that there is a correlation betweendecreasing FDPL and increasing PSNR. Training a model with FDPL results in ahigher average PSRN (30.94), compared to the same model trained with pixel loss(30.59), as measured on the Set5 image dataset. We also show that our methodachieves higher qualitative results, which is the goal of a perceptual lossfunction. However, it is not clear that the improved perceptual quality is dueto the slightly higher PSNR or the perceptual nature of FDPL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×