New AI Capable Of Completing Photos Without Any Experience

In the past few months, new research has been conducted on machine learning algorithms that have the ability to complete and edit single images based on user orders. This is done by uncovering patterns in the image’s data in order to separate the parts desired by the user and the parts that are not. This modern technology has a lot of potential in many different fields, such as for artistic purposes, forensics or observation of nature and animals.

The majority of jobs that machine learning programs undertake require extremely large amounts of data for it to train with. For example, for a machine to recognise a cat, it will need lots of pictures of different types of cats from all different angles. Unfortunately, in many situations such data is not available. Recently, a group of Israeli analysts have been developing something named “deep internal learning,” in which a program deduces the inside composition of a single picture. The researchers’ new project takes further what has previously been achieved by another team called DIP (Deep Image Prior).

The use of new AI technologies such as this could drastically speed up some forensics processes.

The use of new AI technologies such as this could drastically speed up some forensics processes.

This technology employs “deep learning”, a new approach that includes the use of “multi-layer neural networks”. In essence, a Deep Image Prior is taught to replicate a particular image provided by the user. From a view, the process is rather simple. Firstly, the network is given a randomised input and it will output a bunch of pixels. Then, it will compare the produced pixels to the original image. Based on this comparison, the DIP will modify its internal frameworks in order to construct something closer to the original image when the process runs again. The process will re-run thousands of times against the same original image.

The Israeli team have developed this further, and made use of two separate DIPs, which advance independently of each other. As a result, they end up progressing along slightly different patterns that emphasise different aspects of the image. At the end they are combined to create a better image. This could mean filling in blank spots, enhancing an image’ quality, or making it look more realistic. Practical examples include removing watermarks and reflections from glass.

ResearchEdward Bristow