New Deep Learning System to Fix Your Noisy Photos

Everyone of us like to capture the world as it is. But the problem is that we often experience an unwanted degree of noise in the clicked pictures.




With the advancement of technology, a number of software are available which offer variety of image enhancement capabilities. Our smartphone cameras have also matured over the years to click better pictures, but restoring those blurred images has always been a strenuous task. Many apps provide variety of filters and fade-in, fade-out features, but image restoration of grainy pictures is still a challenge. A good news has been heard for all those who love photography, but can’t click clear photos.

American technology firm Nvidia along with researchers from Aalto University and Massachusetts Institute of Technology has proposed an Artificial Intelligence (AI) system which can enhance your grainy photos. Nvidia has mentioned in their official post that their deep learning-based model can learn to fix your noisy images simply by looking at other similar corrupted photos. Noisy images are those which have any type of distortion or grainy element in them.

Clicked pictures can be noisy because of any reasons like low-light photography, magnetic resonance or digital zoom while capturing the pictures. The idea of fixing noisy images has been presented by the researchers recently in International Conference on Machine Learning (ICML) 2018 in Stockholm, Sweden. Their experiment shows that the deep learning algorithm can flawlessly touch-up grainy pictures to automatically remove almost any type of noise.

The AI system based on deep learning considered a hierarchical residual network and a U-Net to restore images for experimentation.

Training the System

  • A dataset consisting of 256×256-pixel crops drawn from 50,000 clear pictures was extracted from ImageNet dataset, MRI scans and computer-generated images.
  • The images were distorted with introduction of randomized noise.
  • The clear image-grainy image pair formed the training set for the system and the system was trained to remove the noise from grainy image to present the clear version of the image.

Example Training Pairs (Source:


  • The performance of the system is tested by introducing various types of synthetic noise like Poisson noise, binomial noise and Random-valued impulse noise to clear images to obtain noisy images.
  • The various types of noise which can be cleared using the above system range from all types of noise removal to even text removal.
  • The system learns by deep learning algorithm to differentiate between the pair of images clear-grainy images and eventually cleaning up and reconstructing the grainy images.
  • Once the system was sufficiently trained, it was tested for generalization on three sets of images to which they had initially added noise.

It was observed that the system could remove noise from the images in just few milliseconds. The results were amazing because the system could produce images which resembled closely to clean images. Once the system is trained and tested, it can bring out improved version of the noisy input images without having to feed any clean examples for reference.

Hence, it is possible to recover images even under complex corruptions without observing clean images and the system performs quite closer to that which uses clean target data. At the same time this is also true that the system cannot learn to pick missing features in the input image.

It is asserted that the system is capable of restoring even a small detail of any photo with significant clarity to match the original image to a great extent.


Performance of AI System (Source:


This AI system can prove to be a boon in areas where the decision making critically depends on the captured images. Some of the fields like space exploration, medical imagery, remote sensing, pattern recognition etc. where it is not possible to click perfect pictures, these intelligent systems add to the precision of decisions taken.

The AI system is not yet available for public use, but in days to come you will certainly not have to delete all your favorite pics which went grainy from your storage to clean the space. Hopefully, you will be able to recover those to get the noise-free version of your images back.

4 Responses to “New Deep Learning System to Fix Your Noisy Photos”

  1. Hellat says:

    You appear to know a lot like this. Like you wrote the book on it or something. This is magnificent blog. An excellent read and I’ll certainly be back.

  2. Mag says:

    Some truly interesting goodies on this website, appreciate it for contribution.

  3. Aaryan says:

    The article is nicely written. Keep up the work!!

  4. Krita says:

    What an interesting read…

Leave a Reply

Your email address will not be published.