Artificial intelligence has been the driving force behind the most important creations of recent months. In the same category falls the newly developed Nvidia artificial intelligence to reconstruct damaged photos. Nvidia software was developed by the researchers of Nvidia and was designed to rebuild photos with missing pixels or damaged pics using partial convolutions.
There are many cool features related to this program, starting with its application itself, removing the unwanted parts of the images, and then the images reconstructing.
In general, almost any photo editing software can implement the process called image inpainting for deleting, for whatever reason, unwanted content, and then allow artificial intelligence to recover the missing parts.
The partial convolutions method
Unlike current systems, this new method of deep learning, called partial convolution, is calculated the outputs for missing pixels depending on the value of the input that must be provided to the neural network.
However, the output for missing pixels is not linked to the input value provided for those pixels.
At the basis of the development work of the team of researchers coordinated by Guilin Liu and based on deep learning techniques of Nvidia, there is a Carthusian activity that is the generation of 55,116 masks made of random strips and drawings of various shapes and sizes.
The neural network is managed with the PyTorch framework with cuDNN acceleration using NVIDIA Tesla V100 GPUs.
Nvidia Impating uses artificial intelligence to reconstruct deteriorated photos
As it has been presented in the press release regarding the program, the model can effectively handle missing parts of any shape, size or deterioration from the edges of the image.
Previous deep learning approaches have focused on rectangular regions around the center of the image and often rely on costly post-processing.
What’s more, the Nvidia artificial intelligence to reconstruct damaged photos, based on partial convolutions, is able to handle completely missing parts of photos of increased sizes.