This is my implementation of the SRGAN Paper (https://arxiv.org/pdf/1609.04802.pdf) using Tensorflow. This document will also detail my understanding to concepts useful for this project.
A resnet is characterized by two important things: Residual Block and Skip Connection, which sort of also go hand in hand. A residual block is a stack of layers set in such a way that the output of a layer is taken and added to another layer deeper in the block. The non-linearity is then applied after adding it together with the output of the corresponding layer in the main path. Each of these Skip Connection Layers along with the input make up a residual block and form the core of making deeper Neural Networks.
The reason why resnets work are because with skip connections and residual blocks you have that it is much easier to learn indenity mapping. The advantage of this is with ReLu activation, if the deep neural network nudges it towards a wrong and negative or zero direction it will be mapped to the indentity of the place from where the skip connection is established. The disadvantage of plain deep neural networks is that for them it is hard to maintain even indentity mapping and thus it resnets help maintain stability with larger amount of layers.
Perceptual Loss Function is formulated as the weighted sum of a content loss and the adversarial loss component scalled down to 10 to -3.
VGG Loss is defined as the euclidean distance between the feature representations of a reconstructed image and reference image. Thus, it is more closer to perceptual similiarity.
This is the negative of log prob that disciminator judges the reconstructed image is a natural high resolution image. We must minimize this to get the proper backpropogation for the Generator.
Dataset is DIV2k(https://data.vision.ee.ethz.ch/cvl/DIV2K/). I don't own the rights to any of these images.