search

hitzhangyu / Self-supervised-Image-Enhancement-Network-Training-With-Low-Light-Images-Only

Self-supervised Image Enhancement Network: Training With Low Light Images Only
137 stars 18 forks source link

HIST EQ #8

Closed tuvovan closed 3 years ago

tuvovan commented 3 years ago

is the hiseq applied to gray image or color image?

hitzhangyu commented 3 years ago

The max channel of RGB on each pixel，equals the V channel of HSV color space

tuvovan commented 3 years ago

so the HSV image is obtained using opencv or any special function?

hitzhangyu commented 3 years ago

I just use the max function in RGB color space to get the max channel (it equals V channel)，and use the retinex model to keep color retinex model = operate on V channel of HSV color，and keep S and V not change (You can find more about this from: https://www.researchgate.net/publication/343878268_Better_Than_Reference_In_Low_Light_Image_Enhancement_Conditional_Re-Enhancement_Networks)

tuvovan commented 3 years ago

so basically you use histogram equalization on V channel of the input image, and try to minimize the loss between it and the max Rgb?

hitzhangyu commented 3 years ago

Yes，the idea is to explicitly control the contrast of the enhanced result，and CNN with retinex model can help to reduce noise and keep more details ，multiple training images can help avoid some problem of histeq ( e.g. dark areas get darker)

tuvovan commented 3 years ago

thanks ! I'm trying to reproduce your work, I will let you know when it is availble!