AI deep learning technology makes defective images clear and becomes a reality

If the original image data is incomplete, then the obtained image is certainly not clear, it is not easy to repair, and now the AI has learned to identify the source data from the broken image and restore it into a clear photo.

Researchers from NVIDIA, MIT, and Aalto University have discovered a way to use AI to fix pixel photos — even if the AI has never seen a clear version of the target photo.

The team used deep learning to teach AI to piece together images, text, or videos to recover images from a variety of noisy piecemeal pieces of information. This means that AI can break through the previous training method, no longer need to rely on seeing the complete photo to restore the complete features of the damaged photo, from now on, AI can reconstruct a clear photo just through the analysis of "damaged data" or more than two damaged photos.

Surprisingly, the ability of AI to get repairs from "bad photos" or "bad data" is sometimes even better than referencing a complete drawing.

AI achieves image repair by utilizing neural networks trained on damaged photos so that it doesn't require clean images, and experiments have shown that nostalgic photos eroded by different kinds of damage can still produce "nearly identical" results of photo quality recovered using normal photo recognition after multiple recognition analyses by AI. One of the most exciting things about the study is that it can significantly reduce the time required for image rendering — the recognition speed will be measured in milliseconds.

The greater value of this deep learning-based approach in the future is that it will be applied to the medical field, helping to further enhance the imaging clarity of MRI (magnetic resonance) scans and the like.