keras-hiNet

Deep Learning Model which is called HiNet for Image Denoising

Made for Dacon in Korea.

Environment

tensorflow-gpu == 1.13.2

keras-radam == 0.15.0

keras == 2.3.1

albumentations == 1.0.3

train environment

Optimization : RAdam

loss : PSNR Loss

Augmentation : Albumentations(Rotate, Flip)

callbacks : Tensorboard, CosineAnnealing, ModelCheckPoint

learning rate : min(1e-7), max(2e-4)