Gone are the days when images with missing parts could never be reconstructed with realistic approach. NVIDIA AI team came up with innovation that can predict and fill missing parts of a photo with realistic results. At least 55,000 incomplete images of different sizes and shapes were generated to train the deep neural network.
The team unveiled its state-of-the-art deep learning model in 2018 that has the ability to edit images through reconstruction of incomplete parts. In order to train their deep neural network, the research team generated over 55,000 masks of different sizes and shapes.
They then used 25,000 masks with holes to test the dataset. For the enhancement of accuracy, the holes were categorized into six divisions based on the input images.
Nvidia’s AI imaging technique can edit and reconstruct images, even if its parts are completely erased. Before Nvidia’s innovation, Adobe Photoshop was widely used to recreate images. Unlike Adobe Photoshop, Nvidia’s technique uses ‘valid’ pixels to determine what details are supposed to be filled.
To enhance its technology, Nvidia used a novel mathematical method to manipulate the AI that compares pixels in the missing parts with other areas, making it appear normal. The only imperfection in Nvidia’s AI-powered imaging tool is that it isn’t so good at creating esthetic reconstructions.
For the purposes of future research and improvement, the team documented its report in a research paper dubbed as “Image Impainting for Irregular Holes using Partial Convolutions.”
Comments
Post a Comment