Blocky video calls could one day be a thing of the past – thanks to an AI that compresses images by simply throwing out large chunks of them, and making up what should be there instead.
Raw digital images and video require large amounts of memory to store, so computers slim them down by chucking away unnecessary information in a process called compression. When this goes well, you don’t notice that anything is missing, because compression algorithms are designed to get rid of things humans don’t pay attention to, like fine colour gradients. When it goes badly, you get a nasty, jarring image.
This is particularly an issue when it comes to video calls, which often take place over bandwidth-limited cellular networks. If you’ve ever been on a call and seen someone’s face crumble into a pixelated mess, you’ve seen compression rear its ugly head.
Advertisement
Sneaky cheat
Now a team at ETH Zurich, Switzerland have come up with an artificially intelligent solution that uses much less memory than other compression algorithms. The team have created a deep neural network that plays a kind of game with itself, with one part of the network attempting to generate a version of the image that contains as little information as possible, while another part judges whether the image is still recognisable as the original.
The generator network has been trained on a database of images to learn the texture and features of various objects – trees, buildings, cars, and so on – and uses this knowledge to cheat. If a picture contains a tree, it can simply conjure up its own version of a tree, meaning the original can be thrown away.
The AI can also be told to preserve certain objects in an image, which could prove particularly useful for video calls, say the researchers. Rather than waste data transmitting your dull bedroom wall, the system could simply keep the interesting pixels of your face and make up the rest.
To test how well their generated images performed, the researchers showed versions of the same image compressed by their AI and a leading standard algorithm to 180 people. At very low data rates, over 80 per cent preferred the AI version.
¸é±ð´Ú±ð°ù±ð²Ô³¦±ð:Ìý
