Image processing arrangements
AssigneeDigimarc Corporation
InventorsRavi K. Sharma, Tomas Filler, Utkarsh Deshmukh, Vahid Sedighianaraki
Life of this patent
Estimating the term…
Abstract
Aspects of the detailed technologies concern training and use of neural networks for fine-grained classification of large numbers of items, e.g., as may be encountered in a supermarket. Mitigating false positive errors is an exemplary area of emphasis. Novel network topologies are also detailed—some employing recognition technologies in addition to neural networks. A great number of other features and arrangements are also detailed.
Claims26
A method of training a neural network, the network being trained to classify objects depicted in images, the method including the acts:
- assigning training images to a plurality of buckets or pools corresponding thereto, said assigning including assigning training images that exemplify a first object class to a first bucket, and assigning training images that exemplify a second object class to a second bucket;
- picking training images from said plurality of buckets to conduct a training cycle for the neural network;
- changing contents of one or more of the plurality of buckets prior to conducting a next training cycle for the neural network;
- wherein the method further includes:
- discovering a hot-spot excerpt within a training image that depicts an object of a particular object class, said hot-spot excerpt being more important to classification of the image by the neural network than another excerpt within said training image;
- creating a new image by overlaying a copy of said hot-spot excerpt on a background image;
- and
- assigning the new image to a bucket that is not associated with said particular object class, and presenting the new image to the neural network during a training cycle;
- wherein presentation of the new image to the neural network has an effect of reducing importance of the hot-spot excerpt in classifying an image that depicts an object of said particular object class.
A method of training a neural network, the network being trained to classify objects depicted in images, the method including the acts:
- assigning training images to a plurality of buckets or pools corresponding thereto, said assigning including assigning training images that exemplify a first object class to a first bucket, and assigning training images that exemplify a second object class to a second bucket;
- picking training images from said plurality of buckets to conduct a training cycle for the neural network;
- changing contents of one or more of the plurality of buckets prior to conducting a next training cycle for the neural network;
- wherein the method further includes:
- discovering a cold-spot excerpt within a first training image that depicts an object of a particular object class, said cold-spot excerpt being less important to classification of the image by the neural network than another excerpt within said training image;
- generating a modified training image by obscuring the cold-spot excerpt from said first training image;
- and
- assigning the modified training image to a bucket that is associated with said particular object class, and presenting the modified image to the neural network during training;
- wherein said modified training image is richer in information useful to the neural network in classifying the training image than said training image prior to modification, and thus serves to help the neural network learn to distinguish depictions of said object from depictions of other objects.
Dependencies are inferred from the claim text.