Nvidia AI can remove noise and artifacts from grainy photos

Shawn Knight

Posts: 12,522   +122
Staff member

Nvidia has developed an impressive deep learning technique capable of automatically removing noise and artifacts from photos. Whereas recent deep learning work in this field has focused on training a neural network with clean and noisy images, Nvidia’s AI can do so without ever being shown a noise-free example.

The new deep learning-based approach, developed alongside researchers from MIT and Aalto University, can remove noise, artifacts and grain after only seeing two “dirty” samples.

Nvidia used Tesla P100 GPUs with the cuDNN-accelerated TensorFlow deep learning framework and trained the system on 50,000 images in the ImageNet validation set. It then validated the neural network on three different datasets.

The results are rather impressive, especially on images corrupted with random text and photos with excessive noise. As anyone that has worked with photo editing software can attest to, however, noise reduction tools and techniques have been around for many years and in the hands of a skilled user, it’s possible to get very good results.

Also, as with traditional techniques, there does appear to be a loss of sharpness when using Nvidia’s AI.

The development team will be presenting its work in an oral presentation at the International Conference on Machine Learning in Stockholm, Sweden this week.

Permalink to story.



Posts: 2,487   +1,759
Yep, you can do pretty much the same thing in photoshop, if you have the patience and understanding
of photoshop. Been using it for 20 years and I still once in a while will run into "I didn't know I could do that"
moments. The AI is nice, but I still prefer to have the control.


Posts: 4,719   +446
I can foresee a lot of watermark removal from photographers images being done with this...
Good stuff, but still, the use of graphic editors is better, as for me. They are more variable. Photoshop, GIMP, online services do the same. That and the video card for this will need a powerful, I think.


Posts: 16,065   +4,868
The text removal ability is the most impressive. Trying to get rid of that much text with a clone stamp tool would take hours, if not days. I'm thinking you could possibly train it to remove street lamp poles, traffic signs, parking meters, people, and even cars or trucks, to clean up architectural images.

(However, presented with samples as small as those in the video, it's almost impossible to know what defects a full size print might contain).

Perhaps adding ray tracing would allow you to alter light quality and angle, better than the tools commonly available now. (Although I confess to only doing a quick read on the concept).

I think there was an article here which claimed that somebody was working on creating a tool to correct different types of out of focus problems in post, which ostensibly, could have even more potential for, "saving that one great shot that would have been totally epic, if it were only in focus".
Last edited: