Developers from Duke University have developed an AI tool that can enhance the resolution of an image and turn into a recognizable image from an extremely blurry one.
Researchers have named their method as PULSE, and it is based on a generative adversarial network. GAN is an old tool in machine learning and consists of two neural networks. The first neural network provides an output and the second network takes that output as an input. The first neural network generates a human face, while the second neural network decides if it is a suitable match to what is required, if not the image is generated in again. AI-based image generated by the first neural network is highly dependent on the images it was trained on.
Up till now, traditional software utilized a different approach while generating high-resolution images. The image is broken into pixels, and another pixel is matched with an existing pixel. If the new pixel sharpens the image, it replaces the existing pixel and the next pixel is inspected. However, there is a limitation to such kind of software, and it could generate a non-realistic looking image while it sharpens it pixel by pixel and not the complete image all at once.
PULSE can enhance the resolution of nearly every image, and not just faces. However, it has been used on headshots so far. This could be of immense help in the fields of medicine and microscopy to astronomy and satellite imagery.
However, this AI tool cannot be used to identified people as it does not produce the exact replica of the image. All it does is add a few extra lines here and there or a few eyelashes to enhance the image resolution. There are chances that the altered image might not represent the same person as it did in its original form. Hence, it could not be used by law enforcement agencies to identify criminals from a CCTV image or some other portrait.
There are several other tools that can convert a low-resolution image to a high-resolution image, but they can sharpen the image up to 8 times only. PULSE is capable of sharpening the image up to 64-times. This means it can convert 16×16 pixel image to 1024×1024 pixel image.