Skip to content

Latest commit

 

History

History
117 lines (66 loc) · 3.83 KB

README.md

File metadata and controls

117 lines (66 loc) · 3.83 KB

A repo for super resolution using basic neural networks. These models can upscale full hd movies to 4k(need to make them faster to run in real time)

Super resoltion is an ill posed problem that doesn't have a unique answer a low resolution image(lr) can correspond to multiple high resolution(hr) images. Specially in over x4 resolution there are lots of lost information, many textures are lost, many different textures can result to the same low resolution pattern.

16x super resolution

comparison 16x

4x super resolution

comparison 4x

low resolution image

low

resolution increased with one layer network

regenerated with one layer

resolution increased with two layer network

regen 2 layers

original high res image

original

car low res image

carlow

car regenerated

regen

car original

original


16x super resolution from 40x40 to 160x160

low res

16low

regenerated

regen1

regenerated 2

regen 2

original

original

comparison

low res image jojo low

upscaled image jojo regen

high res original image jojo regen

low res image jojo low

upscaled image jojo regen

low res image jojo low

upscaled image 16x jojo regen

original image jojo regen

hi res original image jojo regen

supermoviechunks notebook scales a full hd movie to a 4k movie. Here are 3 screens from blade runner model(selected ones with problems try to find them :))

full hd image bladerunner low

4k upscaled image bladerunner high

full hd image bladerunner low

4k upscaled image bladerunner high

full hd image bladerunner low

4k upscaled image bladerunner high

full hd image bladerunner low

4k upscaled image i think this one is better than the original not just the resolution but also the colors bladerunner high