NVIDIA learned how to restore badly damaged photos

NVIDIA научились реставрировать сильно поврежденные фотографии

Using these technologies can improve many lost shots.

The company NVIDIA has developed a technology to recover lost the photos and other images. The feature resembles a tool the Photoshop “Fill by content” (Content-Aware Fill), but far more “realistic”.

The developers used the new technology with elements of artificial intelligence to a remarkably accurate reconstruction of images. A two-minute video presents the results and shows examples of “modification” of the images.

The most amazing new technology NVIDIA is that information to complete a lost parcel is not extracted from the surrounding pixels, as is the case photoshopish “Fill by content”. The NVIDIA “knows” how to look like the object.

For example, if you apply the tool Content-Aware Fill to white “holes” in places where the face should be the eyes – as a result, the gaps filled with fragments of the surrounding image taken from skin, eyebrows, nose and hair.

The algorithms also NVIDIA know that in this area of the face should be the eyes, so it adds eye – self-generating them.

As the tool “Fill by content”, the NVIDIA can be used to restore lost content, and for the removal of defects and correction of failed areas of the image.

The developers of NVIDIA has published an article titled “Retouch of image defects irregular shapes with partial convolution”.

“Our model can effectively cope with holes of every shape, size, location and distance to the image borders,” write the authors. – Previous deep learning approaches have focused on rectangular areas located around the center of the image, and often required extensive post-processing. Our model also correctly handles holes large sizes”.

Researchers have taught your intelligent system with 55 thousand random configurations of holes and cracks, applied to a huge collection of photos. Comparing images with defects and the original image, the neural network was trained to reconstruct the lost pixels. Additional 25 thousand masks were used in the testing stage.

Share Button