This repository contains the code for our work on densoising high resolution images using deep learning paper. Present state-of-the-art methods like BM3D, KSVD and Non-local means do produce high quality denoised results. But when the size of image becomes very high, for ex. 4000 x 80000 pixels, those high quality results come at a cost of high computational time. This time consuming factor serves as a motivation to come up with a model that can provide comparable results, if not better, in much less time. So, I've used a deep learning approach that automatically tries to learn the function that maps a noisy image to its denoised version. I've used thenao as the deep learning framework, and have worked on the publicly available codes provided by the MILA Lab.
Unfortunately, the complete data on which I actually trained this model cannot be released publicly, since I used data that belong to ISRO (images captured by CARTOSAT 2), although I have included snippets of one or two images in the results section, to provide a sense of what the data looks like. But anyone can easily use their own data (black and white noisy and denoised images for now) and train the model accordingly.
python 2.7, theano
In this code files like /AKASHDP_DATA3/ankur/train/A1-NAO-25-FEB-2016-100907-R2-1_b1.rad_boost are binary files which contain the noisy and denoised images during training. The function read_filename reads these binary files and conversts them into stacked numpy arrays suitable for training.
For training your own model, set the desired configuration of the stacked autoencoders and ensure the availability of suitable data in the form of binary files. Run:
python denoise_function.py
Variation of training process with different number of hidden layers
Variation of training process with different sizes of hidden layers
