# Denoise (04-05)-09-22 Log ## Trained Grayscale images # SR GAN* Extra 1 conv layer after first conv layer, **3** Residual Block --- ## Our dataset (the lighthouse@10fps + gauss+s&p+specle) ### Regular Loss Discriminator loss = BCE loss Generator loss = .5 * BCE loss + content loss + pixel loss + smooth loss ***tried with excluding pixel loss and smooth loss tried with only BCE loss*** **Result** At first generator is showing patterns of real image but slowly converges to gray <a href="https://ibb.co/BsfNP75"><img src="https://i.ibb.co/BsfNP75/image.png" alt="image" border="0"></a> # SR GAN** w/ and w/o Extra 1 conv layer after first conv layer, **5 and 8** Residual Block, w/ and w/o upsampling block Weights initialized from normal distribution of mean 0.0 and variation 0.02, mean .5 std .2, mean .9 std 0.02 --- ## Kaggle Dataset ### Wasserstein Loss #### Variation 1: Discriminator is trained N:1 of Generator, guass+s&p+speckle noise added additionally **Discriminator loss** = Mean of discriminator prediction + Gradient Penalty **Generator loss** = .5 * Mean of Discriminator Score + content loss + pixel loss + smooth loss with N=30 ***tried with excluding pixel loss and smooth loss tried with only mean of discriminator score*** **Result**: Just generating gray image and discriminator loss and generator loss is exponentially increasing <a href="https://ibb.co/gM2MXy3"><img src="https://i.ibb.co/gM2MXy3/image.png" alt="image" border="0"></a> #### Variation 2: Discriminator is trained 1:1 of Generator **Discriminator loss** = BCE loss + GP **Generator loss** = .5 * Mean of Discriminator Score + content loss + pixel loss + smooth loss ***tried with excluding pixel loss and smooth loss*** ***tried with only mean of discriminator score*** * in these case generator loss increases slowly, but discriminator loss increases exponentially **Result**: Just generating gray image and discriminator loss and generator loss is exponentially increasing For N = 30 <a href="https://ibb.co/XW8nxLB"><img src="https://i.ibb.co/XW8nxLB/image.png" alt="image" border="0"></a> ### Just BCE Loss for Gen and Disc Discriminator becomes perfect too soon <a href="https://ibb.co/rwgL7JN"><img src="https://i.ibb.co/rwgL7JN/image.png" alt="image" border="0"></a>