US Patent 10885628 Single image completion from retrieved image collections

A method of completing a masked image includes, in part, identifying a multitude of images that are visually similar to the masked image, retrieving a first subset of images from the multitude of images, setting parameters of a neural network to a first set of values in accordance with the data represented by the first retrieved subset, and using the neural network with the first set of parameters to complete the masked image if the neural network having the first set of parameters is determined to meet a threshold value. If the neural network having the first set of parameters is determined not to meet the threshold value, in an iterative manner, another subset of images different from the first subset is retrieved, and the parameters of the neural network are then updated in accordance with the data represented by the other subset until the threshold value is met.