Summary of the technology
a new architecture which breaks the deblurring network into an analysis network which estimates the blur, and a synthesis network that uses this kernel to deblur the image.
Project ID : 47-2020-10908
Description of the technology
Computer Science and Engineering
Neural Networks, Computer Vision
Motion blur due to *camera shake is one of the predominant sources of image degradation in handheld photography. Blind image de-blurring remains a challenging problem for modern artificial neural networks. Unlike other image restoration problems, neural network architectures for de-blurring fall behind the performance of existing de-blurring algorithms in case of uniform and 3D blur models. This gap follows from the diverse and profound effect that the unknown blur-kernel has on the de-blurring operator.
We propose a new architecture which breaks the de-blurring network into two parts: an analysis network (kernel) which estimates the blur; and a synthesis network that uses this kernel to de-blur the image. Unlike existing de-blurring networks, this design allows us to incorporate the blur-kernel expressly into in the network’s training.
Specifically the advantages this technology are:
New cross-correlation layers that allow for better blur estimations, including unique components that allow the estimate blur to control the action of the synthesis de-blurring action.
When we evaluate this approach over established benchmark datasets, we see the ability to achieve state-of-the-art de-blurring accuracy on various tests, as well a major speed up in runtime.
Results on datasets:https://www.cse.huji.ac.il/~raananf/projects/deblurnets/
Implementation within image editing & communication platforms
VP, BUSINESS DEVELOPMENT
HUJI, School of Computer Science and Engineering