In recent years, great strides have been made in the field of artificial intelligence. These improvements are very well applied to photography and image, with increasingly intelligent and efficient compression algorithms. The last of the advances has been demonstrated by NVIDIA, with a deep learning system capable of reconstructing incomplete images.
Although methods like automatic Photoshop refill are able to rebuild chunks of images, the method you use is based on the content you have around. However, NVIDIA’s deep learning method can rebuild things that are not, including eyes or incomplete parts of a face.
Unlike the Photoshop tool, the NVIDIA AI can analyze the image and understand what type of subject or environment it represents. For example, you can detect that it is in front of a young girl or an older person, or a hilly or arid landscape. If you remove an eye, the IA knows, because it was trained with thousands of pictures of people, there should be an eye. So it replaces it with a realistic computer-generated eye.
To find out which parts to rebuild they also used 55,000 random strokes and holes of varying sizes and shapes. They also generated another 25,000 to train, and ordered them into six categories. Throughout the training process they used NVIDIA Tesla V100 and the utility of deep learning PVTorch with cuDNN acceleration.