deepDenoise.txt

NOTES

Reading the paper : Beyond a Gaussian Denoiser:Residual Learning of Deep CNN for image denoising.

GOAL of image denoising :
Derive a clean image x from a noisy image y which follows a degradation model : y = x + v ; v = noise

Aim is to implicitly remove the clean image and learn v . 

This can also be used in image super resolution problem where can see v to be the difference between the 
groud truth high res image and the bicubic upsampling of a low resolution image , similarly JPEG image 
deblocking problem can also be solved by taking v as the difference between the original image and the
compressed image.

Section III
============
Proposed model :
 training a Deep CNN involves 2 tasks: 
    > Network architecture and design
    > Model learning and training data ( adopt Residual-Learning with batch normalization )

A. Network Depth
-----------------
    > filter size : 3 x 3 and remove all pooling layers
    > Receptive field of DnCNN of depth d should be (2d + 1) x (2d + 1)  // why ? what is this ?

B. Network Architecture
-----------------------
    Generally Denoising algorithms tend to try to learn a mapping function such as F(y) = x to learn the 
clean image x , but here we adopt residual learning , we want to map R(y) = v and then we can get the clean
image by doing: x = y - R(y)

LOSS FUNCTION : averaged mean square error between the desired residual image and estimated ones from noisy input

Given we have a net of depth D we are gonna have 3 types of layers :
    1) Conv + Relu : 
        > This is the first layer 
        > has 64 filters of (3 x 3 x c)

    2) Conv + BN + Relu :
        > from layer 2 till layer (D-1)
        > 64 filters of size (3 x 3)
        > do Batch Norm after Conv and before ReLU
    
    3) Conv
        > Last Later
        > 'c' number of filters of size (3 x 3 x 64) is used to recostruct the output

    * Reducing Boundry artifacts
        > zero pad intermediate layers so as to retain the original size as input image

EXPERIMENTAL RESULTS AS MENTIONED BY THE AUTHORS OF THE PAPER
> Training and Testing Data:
    TRAIN
        DnCNN-S (S for specific noise level)
        > 400 images of 180 x 180 for training
        > considered noise levels of sigma = 15,25,50
        > patch size = (40 x 40)
        > crop = (128 x 1)
        > 600 patches to train the model

        DnCNN-B (B for blind noise level)
        > noise levels sigma = [0,55]
        > patch size = (50 x 50)
        > crop = (128 x 3000)
    
    > data-augmentation : rotate/flip pairs within mini-batch
    
    > depth = 17(DnCNN-S) , 20 (DnCNN-B)

    > SGD with decay = 0.0001 , momentum = 0.9 , mini-batch size =128

    >epoch = 50

    > lr = 1e-1 to 1e-4 over the 50 epoch