search

Saafke / EDSR_Tensorflow

TensorFlow implementation of 'Enhanced Deep Residual Networks for Single Image Super-Resolution'.
Apache License 2.0
233 stars 72 forks source link

About the mean rgb of DIV2K dataset #8

Closed johnniewalker3 closed 3 years ago

johnniewalker3 commented 3 years ago

Hello my friend, Your project is so helpful.i have got many ideas from the way you have written it.I would want to ask you some questions.When we find the mean rgb value of div2k dataset , are we using the Low resolution images of train dataset(800 images) or are we using the low resolution images of train,validation and testing to find the mean rgb value(all the low resolution images of the dataset)?This mean value is this value we must subtract from our data when we pass them in the network

Saafke commented 3 years ago

Hi Johnnie,

When you calculate the mean you always only use the training dataset (otherwise it's 'cheating').

johnniewalker3 commented 3 years ago

Thanks for the soon answer.Also I noticed you are using the mean value of high resolution training dataset.we can use either low resolution or high resolution training dataset to calculate mean RGB value because their difference is very small.I had made a program to check this before continue to the next step which was the creation of the network and after this the training step