Neural networks have become an increasingly promising way to automate the restoration of damaged and/or low-resolution images. Some of these systems are publicly available, such as Let`s Enhance, while others are being demonstrated through published research projects. The latest example of such technology is called Deep Image Prior, but it deviates from the norm. Rather than being pre-trained with a data set, Deep Image Prior recreates an image using the image itself as a guide.
In a paper detailing how Deep Image Prior works, researchers explain:
In this work, we show that, contrary to expectations, a great deal of image statistics are captured by the structure of a convolutional image generator rather than by any learned capability. This is particularly true for the statistics required to solve various image restoration problems, where the image prior is required to integrate information lost in the degradation processes.
In order to demonstrate this, the team used untrained deep convolutional neural networks with a single degraded image. The paper describes this as a `very simple formulation` that is capable of in-painting, denoising, and increasing the resolution of the degraded image without needing to be `trained` with thousands of images beforehand. In essence, it`s like the more advanced, AI-powered version of Content Aware Fill called `Deep Fill` that Adobe showed off at MAX.
The team provided samples of various images processed by Deep Image Prior, including ones where JPEG compression artifacts were removed, image noise was removed, a low-resolution was upscaled to a larger resolution, and missing portions of an image were replaced via in-painting. The system is also capable of removing text from images, raising copyright theft concerns.
Though Deep Image Prior isn`t perfect, the examples provided are impressive, and hint at a future in which such tools may be commonly available to consumers.
Whether that is a good thing remains ...
|
