Deep Reflection Prior

single branch FCN

• assumption: background are the same
• I -> B + R
  • recons loss, cross recons loss, floor/ceiling rejection loss
    • how could such a loss work?
• latent code of B and R feed into network -> best preformance at first norm layer