EP0482427A2
Expired / Lapsed
EP
⏱️ 0.0 years remaining
Self adaptive hierarchical target indentification and recognition neural network
Assignee: Raytheon Co
•
Filing: 1991-10-09
•
Granted: —
•
Expires: 2011-10-09
Abstract
Abstract A self adaptive hierarchical target identification neural network pattern recognition system (10) is constructed utilizing four basic modules. The first module is a segmenter and preprocessor (14) which accepts gray level image data (12) and is based on the Boundary Contour System neural network. The segmenter and preprocessor (14) output is fed to a feature extractor (16) which comprises a first layer of a neocognitron. The feature extractor (16) output is fed to a pattern recognizer (18) which comprises two layers of the neocognitron. The pattern recognizer (18) produces as output a real valued vector representation which encodes the object to be identified. This vector representation is fed to a classifier (20) which comprises a back propagation neural network. The pattern recognition system (10) can classify large numbers of objects from raw sensor data and is relatively translation, rotation and scale invariant
Inventors
- 👤 C. A. Daniell
- 👤 W. A. Tackett
- 👤 J. A. Alves
- 👤 J. A. Burman
- 👤 K. B. Johnson
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-11-07 11:43
CPC Classifications
G; G06; G06N; G06N3/00; G06N3/02; G06N3/04; G06N3/0463; G06V; G06V10/00; G06V10/40; G06V10/44; G06V10/443; G06V10/449; G06V10/451; G06V10/454
Claims Analysis
Total Claims: 13
Claim 1
Independent Claim
125 words
A system for locating and classifying a target image contained in an input image (12), characterized by: - a segmenter (14) for producing as output a binary image (34) representing the input image (12); - a feature extractor (16) coupled to said segmenter (14) for receiving said output from said segmenter (14) and for generating output responses that characterize elemental features in said input image (12); - a pattern recognizer (18) coupled to said feature extractor output for combining a plurality of said elemental feature responses to form new parallel output responses that characterized complex combinations of the elemental features; and - a classifier (20) coupled to said pattern recognizer output for producing an output that corresponds to a predetermined classification of said target image.
Claim 2
Dependent Claim
17 words
The system of claim 1, characterized in that said input image (12) comprises raw sensor imagery data.
Claim 3
Dependent Claim
14 words
The system of claim 2, characterized in that said data comprises infrared sensor data.
Claim 4
Independent Claim
119 words
A system for recognizing patterns in input data, characterized by: - a segmenter (14) for producing as output a binary signal representing said input pattern; - a feature extractor (16) coupled to said segmenter (14) for receiving said parallel output from said segmenter (14) and for generating output responses that characterize elemental features in said input data; - a pattern recognizer (18) coupled to said feature extractor output for combining a plurality of said elemental feature responses to form new parallel output responses that characterize complex combinations of the elemental features; and - a classifier (20) coupled to said pattern recognizer output for producing an output that corresponds to a predetermined classification of said patterns in said input data.
Claim 5
Independent Claim
48 words
The system of any of claims 1 through 4, characterized in that said segmenter (14) comprises a Boundary Contour System (22) for separating said image (12) into a plurality of gradient images at equally spaced orientations, whereby said target is localized and separated within said input image (12).
Claim 6
Dependent Claim
36 words
The system of claim 5, characterized in that said Boundary Contour System (22) includes a system of Shunting Membrane Equation neurons to form large responses where gradient boundaries exist in the input image or date, respectively.
Claim 7
Dependent Claim
101 words
The system of claim 5 or 6, characterized in that said Boundary Contour System (22) comprises: - means (24) for producing a gradient phase field using oriented masks; - means (26) for generating a spatial gradient using center surround; - means (28) for generating local competition across orientations; and - a feedback mechanism (30, 32) including - means (30) for generating global coopertion among orientations whereby missing or weak edge information in said image is filled in, and - means (32) for generating a spatial gradient using center surround for removing blurring caused by the means (30) for generating global cooperation.
Claim 8
Independent Claim
25 words
The system of any of claims 1 through 7, characterized in that said feature extractor (16) and said pattern recognizer (18) together comprise a neocognitron.
Claim 9
Dependent Claim
94 words
The system of claim 8, characterized in that said feature extractor (16) includes a first neocognitron layer (40, 42) and said pattern recognizer (18) comprises second and third neocognitron layers (46, 48, 50, 52), each of said first, second and third layers (40, 42, 46, 48, 50, 52) comprising two sub-layers (40/42, 46/48, 50/52), each sub-layer (40/42, 46/48, 50/52) having a two-dimensional array (36) of neurons, each of said neocognitron layers (40, 42, 46, 48, 50, 52) being connected by convolutional masks (38, 44) consisting of weighted connections between selected ones of said neurons.
Claim 10
Dependent Claim
19 words
The system of claim 9, characterized by means for adjusting the strength of said weighted connections by unsupervised training.
Claim 11
Independent Claim
19 words
The system of any of claims 8 through 10, characterized in that said neocognitron comprises a selective attention neocognitron.
Claim 12
Independent Claim
24 words
The system of any of claims 1 through 11, characterized in that said pattern recognizer (18) produces a vector representation of said target image.
Claim 13
Dependent Claim
135 words
A missile guidance system for locating and classifying a target image contained in an input image (12), characterized by: - a Boundary Contour System (22) coupled to said input image (12) for producing as output a plurality of gradient orientation response images in parallel; - a neocognitron comprising three layers (40/42, 46/48, 50/52), wherein said first layer (40/42) is coupled to said Boundary Contour System (22) and comprises a feature extractor (16), and the second and third layers (46/48, 50/52) comprise a pattern recognizer (18) coupled to said feature extractor (16) for generating parallel output response that characterize complex combinations of elemental features in said input image (12); and - a back propagation neuron network coupled to said pattern recognizer (18) for producing an output that corresponds to a predetermined classification of said target image.
References Cited (Backward Citations)
No citation records indexed.
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.