/Deep-Image-Prior

PyTorch implementation of the CVPR 2018 paper Deep Image Prior by Dmitry Ulyanov et. al.

Primary LanguageJupyter Notebook

Deep Image Prior

PyTorch implementation of the CVPR 2018 paper Deep Image Prior by Dmitry Ulyanov et. al. This codebase is a part of final project for CS 663: Fundamentals of Digital Image Processing at IITB. The team members are:

We use the method described in the paper for image inpainting. We randomly delete some pixels of a 512*512 RGB image and then try to fill them using the said technique.

Results

50% pixels removed

The images used are the standard Lena and Barbara test images. The input and resultant images are as shown.The progress in the initial iterations is shown on the rightmost column.

80% pixels removed

Training

This are the loss curves we obtained during the training on 50% and 80% images respectively. We can observe that the reconstruction loss decreases to negligible value within very few iterations.

References

  1. Dmitry Ulyanov et. al Deep Image Prior [arxiv]
  2. Test Images Public-Domain Test Images for Homeworks and Projects