Method and apparatus for training a neural network to detect objects in an image

• US 6,324,532 B1
• Filed: 12/15/1999
• Issued: 11/27/2001
• Est. Priority Date: 02/07/1997
• Status: Expired due to Term

• Current Assignee: SRI International, Inc.
• Original Assignee: Sarnoff Corporation (SRI International, Inc.)
• Inventors: Sajda, Paul, Spence, Clay Douglas
• Primary Examiner(s): Davis, George B.
• Application Number:
  US09/464,506
• Time in Patent Office: 713 Days
• Field of Search: 706/27, 706/19
• US Class Current: 706/27
• CPC Class Codes: G06F 18/24133 Distances to prototypes
• CPC Class Codes: G06F 18/24155 Bayesian classification
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/082 modifying the architecture, e.g. adding, deleting or silencing nodes or connections
• CPC Class Codes: G06V 10/255 Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
• CPC Class Codes: G06V 10/82 using neural networks

Abstract

A signal processing apparatus and concomitant method for learning and integrating features from multiple resolutions for detecting and/or classifying objects. The signal processing apparatus comprises a hierarchical pyramid of neural networks (HPNN) having a “fine-to-coarse” structure or a combination of the “fine-to-coarse” and the “coarse-to-fine” structures.

Claims

1. An architecture of a plurality of neural networks for processing an input signal that is decomposed into a plurality of resolution levels, said architecture comprising:

a first neural network for receiving input from a fine resolution level of said input signal;

a second neural network for receiving input from a coarse resolution level of said input signal; and

a third neural network having an input layer for receiving input from a resolution level that is coarser than said fine resolution level and finer than said coarse resolution level of said input signal and for receiving inputs from an output layer of said first neural network and from an output layer of said second neural network.

2. An architecture of a plurality of neural networks for processing an input signal that is decomposed into a plurality of resolution levels, said architecture comprising:

a first neural network for receiving input from a fine resolution level of said input signal;

a second neural network for receiving input from a first intermediate resolution level that is coarser than said fine resolution level and for receiving input from said first neural network;

a third neural network for receiving input from said first intermediate resolution level and for receiving input from a neural network that is receiving input from a resolution level that is coarser than said first intermediate resolution level;

a fourth neural network for receiving inputs from a coarse resolution level of said input signal;

a fifth neural network for receiving input from a second intermediate resolution level that is finer than said coarse resolution level and for receiving input from said fourth neural network; and

a sixth neural network for receiving input from said second intermediate resolution level and for receiving input from a neural network that is receiving input from a resolution level that is finer than said second intermediate resolution level.

3. The architecture of a plurality of neural networks of claim 2, further comprising:

a seventh neural network for receiving input from said second and said third neural networks; and

an eighth neural network for receiving input from said fifth and said sixth neural networks.