This work is about removing noise/haze from noisy/hazy images to obtain clearer images using stacked autoencoders.

The work is a result of participation in (New Trends in Image Restoration and Enhancement workshop) NTIRE 2020 Real Image Denoising Challenge - Track2 – sRGB, in conjunction with (Conference on Computer Vision and Pattern Recognition) CVPR 2020, Seattle, US.

Participation in NTIRE 2020 Denoising Challenge led my team secure 16th position among 250+ registrants. Publication Link

This work also discusses implementation on Image Dehazing as an extension to the original work on Image Denoising. Our results from the Pix2Pix model position among the Top-16 out of 270+ registrants for the NTIRE 2019 Image Dehazing Challenge. Report

Further work was done post the competition. The summary of results is as follows:

From the results, we can conclude that our basic PyTorch implementation has been able to perform quite well in general, and compared to two state-of-the-art approaches for Image Translation tasks - CycleGAN and Pix2Pix.

• PyTorch
• OpenCV

• To install PyTorch, see installation instructions on the PyTorch website.
• To install OpenCV: pip install opencv-python

• This work makes use of the Smartphone Image Denoising Dataset SIDD. The data was made available on the Codalab website where the competition was hosted.

• This work makes use of Dense-Haze dataset for the purpose of Image Dehazing.

The data in original forms are extremely high resolution images. Due to lack of computing resources, for training purposes, the images have been down-sized to 256 pixels.

• Use the denoising_make_data.py script to process images and convert them to a numPy array (.npy) file.

• Use the dehazing_make_data.py script to process images and convert them to a numPy array (.npy) file.

• Denoising Autoencoder
• CycleGAN & Pix2Pix (Official)