Awesome

DySample Unet

This project is an implementation of Unet with the DySample upsampler used in the decoding block. According to tests in the SR field as a GAN, this approach to semantic segmentation increases stability during cold start and the standard training process. All tests were conducted using NeoSR and the SPAN SR architecture.

blue - Unet | green - DUnet

xychart-beta
    title "Zero start: Unet vs DUnet"
    x-axis [5k, 10k, 15k, 20k, 25k, 30k, 35k, 40k, 45k, 50k, 55k, 60k, 65k, 70k, 75k, 80k, 85k]
    y-axis "SSIM (higher is better)"
    line [0.5965887904167175, 0.6561548709869385, 0.6582467555999756, 0.3635033071041107, 0.44282767176628113, 0.6836576461791992, 0.6523440480232239, 0.6668142676353455, 0.7022196054458618, 0.6715793609619141, 0.3366641104221344, 0.7047410011291504, 0.2585214674472809, 0.7050371766090393, 0.41283008456230164, 0.6888425946235657, 0.5920573472976685]
    line [0.6593372225761414, 0.6420486569404602, 0.6501752734184265, 0.6584635972976685, 0.6524503231048584, 0.6740251779556274, 0.6513743996620178, 0.6689457297325134, 0.6556380391120911, 0.6701475381851196, 0.6934555768966675, 0.6391361355781555, 0.6657055616378784, 0.6925287842750549, 0.6902399063110352, 0.685698390007019, 0.6594018936157227]

blue - EA2FPN | green - DUnet

xychart-beta
    title "Real start: EA2FPN vs DUnet"
    x-axis [5k, 10k, 15k, 20k, 25k, 30k, 35k, 40k, 45k, 50k]
    y-axis "SSIM (higher is better)"
    line [0.6245308518409729, 0.6849638223648071, 0.6963799595832825, 0.6750068068504333, 0.7084226608276367, 0.6957032084465027, 0.7006552815437317, 0.6688494682312012, 0.6998009085655212, 0.6535440683364868]
    line [0.7206200361251831, 0.7372534275054932, 0.7425410747528076, 0.7406359314918518, 0.7393627762794495, 0.7396222352981567, 0.7326762080192566, 0.7352635264396667, 0.7331317067146301, 0.7371227741241455]