The proverbial “picture is worth a thousand words” has been reimagined in the age of the internet. Folks using Facebook, Snapchat, Instagrams are uploading tons of pictures to convey something and some of us are simply frenzied uploaders. What most of us don’t realize is that stuff uploaded on the internet takes space and we need servers for that, in fact, the very article I am penning down will be stored on the server when you are reading this. RAISR (Rapid and Accurate Image Super Resolution) is a new technology that was introduced in November and was capable of producing better quality versions of low-resolution images which would eventually lead to 75 percent reduction in bandwidth.
RAISR is touted to be better than the current super-resolution methods and is also nearly 10-100 times faster which is a desirable trait when it comes to mobile usage. Upsampling is a process of producing an image of larger size with more pixels from the low-resolution ones. This is usually achieved by filling in new pixels by using fixed combinations of the nearing pixel values but these methods fail to bring out the vivid details in the pictures and the final images would still end up looking blurry and lacking in details.
Google uses the RAISR in conjunction with machine learning trained by pairs of images, one low quality, one high in order to find filters that when applied would recreate the quality that is comparable to the original. The training can be done in two methods the first one is the “direct” method wherein the filters are derived directly from the pair of low and high-resolution image pairs. The second method, however, uses the usual upscaling method and then learns the filter from the upscales and the high-resolution image pairs thus achieving a better sampling.
Both the methods are trained by using the edge features which is usually found in patches of images including brightness/color gradients, flat/textures regions and coherence. That said the problem here is that RAISR needs samples to learn from and in absence of any it might fail in working as intended. As of now Google has rolled out the RAISR for high-resolution images on Google+ and are reducing users bandwidth by about a third and it goes without saying that Google will introduce the same across several other services, Google Photos perhaps.