Video deblurring using neural networks

A system comprising:

• a blur simulator means for combining groups of sequential source images from a source video together to generate a sequence of simulated blurred images;
• a network trainer means for training a neural network by:
• providing training images to a neural network comprising a central training image from the sequence of simulated blurred images, an earlier training that precedes the central training image in the sequence and a later training image that follows the central training image in the sequence;
• evaluating similarity between a training output image generated by the neural network from the training images and at least one reference image;
• and
• modifying the neural network based on the evaluated similarity;
• and
• a network manager means for causing the trained neural network to generate a deblurred output image from a blurry input image.

One or more computer-readable media having a plurality of executable instructions embodied thereon, which, when executed by one or more processors, cause the one or more processors to perform a method for deblurring images, the method comprising:

• training a neural network using a set of iterations, said training comprising for each iteration:
• selecting, from a sequence of blurred images associated with a simulated blurred video, a central training image and a set of neighboring training images surrounding the central training image in the sequence, wherein the central training image is selected based on a blur level of the central training image;
• generating, by the neural network, a training output image based on the central training image and the set of neighboring training images;
• and
• adjusting the neural network based on a comparison between the training output image and a reference image.