/DGCN

Deep Graph Convolutional Network for image restoration.

Primary LanguageTeX

DGCN: Deep Graph Convolutional Network for image denoising.

This repository contains a PyTorch implementation of the paper Deep Graph Convoluitonal Image Denoising.

File Descriptions:

  • analyze.py: contains functions for obtaining training curves, test-set performance, etc. given an argument file (i.e. $ python3 analyze.py path/to/args.json)
  • args.json: sample arguments file used to interface with model configurations and checkpoints.
  • data.py: defines datasets and data-loading functions.
  • knn.py: defines functions for peforming K-nearest neighbors with local-masks.
  • net.py: defines GCDN network and submodules.
  • train.py: used like analyze.py, initializes and fits a model to training data given an arguments file.
  • utils.py: defines functions for data pre/post processing, indexing, etc.
  • visual.py: visualization tools written with matplotlib package, such as an interactive receptive field viewer.

Further Improvements:

  • Faster nearest neighbors computations, perhaps with packages such as FAISS.