Image processing arrangements
AssigneeDigimarc Corporation
InventorsRavi K. Sharma, Tomas Filler, Utkarsh Deshmukh, William Y. Conwell
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.
Claims25
An image processing system including one or more cameras, a point of sale terminal, and a neural network coupled to receive imagery of items from the one or more cameras, and being configured to provide output data identifying types of said items to the point of sale terminal, the neural network comprising:
- one or more input stages coupled to the one or more cameras to receive image data;
- one or more intermediate stages, cascaded following said one or more input stages, each of said intermediate stages having an input coupled to an output of a previous stage;
- and at least first and second concluding stages, each having an input coupled to an output of a common intermediate stage, said common intermediate stage being configured to provide a common set of plural binary output data to each of said plural concluding stages;
- the first concluding stage having N output neurons and the second concluding stage having M output neurons, where M<N, and wherein output neurons of both the first and second concluding stages are configured to provide output data of the same variety, namely output data identifying said types of items.
A method to identify types of items depicted in imagery captured by one or more cameras at a store checkout, the imagery depicting items in a set of items stocked by the store, the method comprising the acts:
- providing a neural network including one or more input stages coupled to the one or more cameras to receive image data;
- one or more intermediate stages, cascaded following said one or more input stages, each of said intermediate stages having an input coupled to an output of a previous stage;
- and at least first and second concluding stages, each having an input coupled to an output of a common intermediate stage, said common intermediate stage providing a common set of plural binary output data to each of said plural concluding stages, each of the first and second concluding stages including at least two layers of neurons, the first concluding stage having N output neurons and the second concluding stage having M output neurons, wherein M<N, wherein output neurons of both the first and second concluding stages are configured to provide output data of the same variety, namely output data identifying said types of items, the method further including training the network to respond to image data of items in a first subset of items by firing output neurons in the first concluding stage;
- and training the network to respond to image data of items in a second subset of items by firing output neurons in the second concluding stage;
- the method further including retraining the network when items are added to said set of items stocked by the store, said retraining comprising retraining the second concluding stage but not retraining the first concluding stage.
A method comprising the acts:
- providing a neural network including one or more input stages coupled to one or more cameras to receive image data;
- one or more intermediate stages, cascaded following said one or more input stages, each of said intermediate stages having an input coupled to an output of a previous stage;
- and at least first and second concluding stages, each having an input coupled to an output of an intermediate stage, each of the first and second concluding stages including at least two layers of neurons, the first concluding stage having N output neurons and the second concluding stage having M output neurons, where M<N, wherein output neurons of both the first and second concluding stages are configured to provide output data of the same variety, namely output data identifying grocery items by type;
- training the network to respond to images depicting first and second grocery items by firing first and second output neurons, respectively, in a single one of said concluding stages;
- and
- retraining the network when items of new types are added to a grocery's offerings, said retraining comprising selectively retraining the second concluding stage but not the first concluding stage.
Dependencies are inferred from the claim text.