High frequency noise streamlines server-based image processing.
Processing high-resolution images on a cellphone is computationally intensive and quickly drains batteries.
One way of getting around this problem is to upload images to a central server for processing and then download them afterward. Unfortunately, the gain in computational power tends to be offset by the extra bandwidth necessary to upload and download high-resolution images.
However, a new system for image processing ameliorates this issue by sending the server highly compressed images and having the server send back simple instructions for modifying the original image.
Researchers from MIT, Stanford and Adobe reported that in their experiments the system reduced the bandwidth of server-based image processing by as much as 98.5 percent. Power-consumption was reduced by 50-85 percent and processing time by 50-70 percent.
Compressing and Manipulating the Image
The system begins by sending a very low quality JPEG to the server. It then introduces high-frequency noise into the image, which effectively increases its resolution. Adding small, random and local variations in pixel color prevents the system from relying too heavily on color consistency when determining how to characterize image transformations.
In order to manipulate the image, the system breaks it into chunks and uses a machine-learning algorithm to characterize the effects of the manipulation according to a few basic parameters—mostly pertaining to variations in luminance or brightness. The server then sends those parameters as a small set of numbers back to the user’s device.
Less Bandwidth, More Computation
The set of numbers sent back to the phone describes modifications for it to make on its local, high-resolution copy of the image.
To the naked eye, the resulting image is virtually indistinguishable from one created through direct manipulation.
Since the high-resolution image is neither uploaded nor downloaded from the phone, bandwidth consumption using this system is a mere 1-2 percent of what it would have been otherwise.
The researchers acknowledged that their system requires some extra computation on phones, since it still needs to manipulate the image according to the parameters sent by the server. However, they contend that it still uses less time and energy than uploading and downloading high-resolution files.
Michaël Gharbi, who presented at the system at the Siggraph Asia conference, believes it will become more useful as image-processing algorithms become more sophisticated.
“We see more and more new algorithms that leverage large databases to make a decision on the pixel,” Gharbi says. “These kinds of algorithm don’t do a very complex transform if you go to a local scale on the image, but they still require a lot of computation and access to the data. So that’s the kind of operation you would need to do on the cloud.”
For more information, view the conference proceedings.