Neural network image processing system

• US 4,941,122 A
• Filed: 01/12/1989
• Issued: 07/10/1990
• Est. Priority Date: 01/12/1989
• Status: Expired due to Term

• Current Assignee: Bantec Incorporated
• Original Assignee: Recognition Equipment Incorp
• Inventors: Weideman, William E.
• Primary Examiner(s): Lastova, John R.
• Application Number:
  US07/296,520
• Time in Patent Office: 544 Days
• Field of Search: 364/800, 364/807, 364/602, 364/300, 364/131, 364/200, 364/900, 364/513, 307/201, 382/14, 382/15
• US Class Current: 708/801
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06V 10/774 Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting

Abstract

A neural-simulating system for an image processing system includes a plurality of networks arranged in a plurality of layers, the output signals of ones of the layers provide input signals to the others of the layers. Each of the plurality of layers include a plurality of neurons operating in parallel on the input signals to the layers. The plurality of neurons within a layer are arranged in groups. Each of the neurons within a group operates in parallel on the input signals. Each neuron within a group of neurons operates to extract a specific feature of an area of the image being processed. Each of the neurons derives output signals from the input signals representing the relative weight of the input signal applied thereto based upon a continuously differential transfer function for each function.

Claims

1. a neural-simulating system for processing an input stimulus, comprising:

a plurality of layers, each layer receiving layer input signals and generating layer output signals, said layer input signals including signals from the input stimulus and ones of said layer output signals from only previous layers within said plurality of layers;

each of said plurality of layers including a plurality of neurons operating in parallel on said layer input signals applied to said plurality of layers;

each of said neurons deriving neuron output signals from a continuously differentiable transfer function for each of said neurons based upon a combination of sets of weights associated within said neurons and said layer input signals;

said plurality of neurons with a layer being arranged in groups of neurons operating in parallel on said layer input signals; and

weight constraining means associated with each of said group of neurons for causing each of said groups of neurons to extract a different feature from said layer input signals and for further causing each neuron with a group of neurons to operate on a different portion of said layer input signals.

2. The neural-simulating system of claim 1 and further including:

adaptive network means associated with each neuron for generating a weight correction signal based upon gradient estimate signals and convergence factors signals of each neuron and for processing said weight correction signals to thereby modify said weights associated with each neuron; and

weight correction signal constraining means associated with each neuron within said groups of neurons for causing each neuron with a group of said neurons to extract the same feature from said layer input signals and for further causing each neuron with a group of neurons to operate on a different portion of said layer input signals.

3. A neural-simulating system for an image processing system, comprising:

a plurality of layers, each layer receiving layer input signals and generating layer output signals, said layer input signals including signals from the image and ones of said layer output signals from only previous layers within said plurality of layers;

each of said plurality of layers including a plurality of neurons operating in parallel on said layer input signals applied to said plurality of layers;

each of said neurons deriving neuron output signals from a continuously differentiable transfer function for each of said neurons based upon a combination of sets of weights associated with said neurons and said layer input signals;

said plurality of neurons within a layer being arranged in groups of neurons operating in parallel on said layer input signals; and

weight constraining means associated with each of said groups of neurons for causing each of said groups of neurons to extract a different feature from said layer input signals and for further causing each neuron with a group of neurons to operate on a different portion of said layer input signals.

4. The neutral-simulating system of claim 3 and further including:

adaptive network means associated with each neuron for generating a weight correction signal based upon gradient estimate signals and convergence factors signals of each neuron and for processing said weight correction signals to thereby modify said weights associated with each neuron; and

weight correction signal constraining means associated with each neuron within one of said groups of neurons for causing each neuron with a group of said neurons to extract the same feature from said layer input signals and for further causing each neuron with a group of neurons to operate on a different portion of said layer input signals.

5. A neutral-simulating system which is trained on a set of input patterns to the system for subsequent identification of unknown patterns, comprising:

clustering means for extracting a training set of patterns representing samples of the input patterns and for updating the training set from the input patterns to the system;

a plurality of layers, each layer receiving layer input signals and generating layer output signals, said layer input signals including signals from the training set and ones of said layer output signals from only previous layers within said plurality of layers;

each of said plurality of layers including a plurality of neurons operating in parallel on said layer input signals applied to said plurality of layers;

each of said neurons deriving neuron output signals from a continuously differentiable transfer function for each of said neurons based upon a combination of sets of weights associated with said neurons and said layer input signals; and

adaptive network means associated with each neuron for generating a weight correction signal based upon gradient estimate signals and convergence factors signals of each neuron and for processing said weight correction signals to thereby modify said weights associated with each neuron.

6. The neural-simulating system of claim 5 and further including:

said plurality of neurons within a layer being arranged in groups of neurons operating in parallel on said layer input signals;

weight constraining means associated with each of said groups of neurons for causing each of said groups of neurons to extract a different feature from said layer input signals and for further causing each neuron with a group of neurons to operate on a different portion of said layer input signals; and

weight correction signal constraining means associated with each neuron within one of said groups of neurons for causing each neuron with a group of said neurons to extract the same feature from said layer input signals and for further causing each neuron with a group of neurons to operate on a different portion of said layer input signals.