This project builds a combined model by merging a trained Pixel loss model based on EDSR with a GAN model based on SRGAN. The project's purpose was to evaluate if this approach can produce better results.
The notebook contains all the code created and used in the project; it consists of Three main parts.
Datasets used in this project are included in "Original_Datasets" folder.
DIV2K dataset used for creating the training and validation datasets; however, only the 'HR' and 'X2 bicubic'subsets were used.
N.B DIV2K datasets were not included due to their size, they can be downloaded using the code in Jupyter Notebook.
Four datasets -Set5, Set14, BSDS100, Urban100- were used for testing.
Three models were built and trained in this project, all trained version are saved in Models folder.
-
Base Model: 2X_base_model.h5, it can be used to generate High-resolution images.
-
GAN Generator: 2X_GAN_gen_model.h5 it can be used to generate High-resolution images.
-
GAN Discriminator: 2X_GAN_dis_model.h5.
Two evaluation test were conducted between the base model, GAN model, and bicubic interpolation.
-
The average PSNR and SSIM were caluclated and compared.
-
Some images were slected to visuale check and compare the results.