sovit-123 / image-deblurring-using-deep-learning

PyTorch implementation of image deblurring using deep learning. Use a simple convolutional autoencoder neural network to deblur Gaussian blurred images.

README

• First of all, you can find the dataset on Kaggle:

  • Dataset => https://www.kaggle.com/kwentar/blur-dataset.

• Get the dataset and extract it inside the input folder. Following is the directory structure for the project:

  ├───input
  │   ├───defocused_blurred
  │   ├───gaussian_blurred
  │   ├───motion_blurred
  │   └───sharp
  ├───outputs
  │   └───saved_images
  └───src
  

Steps to Execute

• I have not used the blurred images that are given in the original dataset for image deblurring. They are spatially variant due to motion blurring and defocus-blurring. I have added Gaussian blurring to the images using the add_guassian_blur.py script inside the src folder. Then I have used these images for deblurring.
• The following is the order of execution:
  1. add_gaussian_blur.py
  2. deblur_ae.py
• Note: Execute all the scripts while being within the src folder inside the terminal.

Some Results

• Loss Plot

  

• Blurred Image

  

  Final Deblurred Image

  

Future Work

• To deblur the spatially variant images inside the defocused_blurred and motion_blurred folders.
• Add more and better models to models.py script.
• Any useful contribution to the project is highly appreciated.

References

• Paper: Image Deblurring with BlurredNoisy Image Pairs, Lu Yuan, Jian Sun, Long Quan, Heung-Yeung Shum.
• Image super-resolution as sparse representation of raw image patches, Jianchao Yang†, John Wright‡, Yi Ma‡, Thomas Huang†.
• mage Deblurring and Super-Resolution Using Deep Convolutional Neural Networks](https://www.researchgate.net/publication/328985265_Image_Deblurring_and_Super-Resolution_Using_Deep_Convolutional_Neural_Networks), Fatma Albluwi, Vladimir A. Krylov & Rozenn Dahyot.
  • GitHub Code.