Method and apparatus for training a neural network to learn hierarchical representations of objects and to detect and classify objects with uncertain training data

• US 6,018,728 A
• Filed: 02/07/1997
• Issued: 01/25/2000
• Est. Priority Date: 02/09/1996
• Status: Expired due to Term

• Current Assignee: SRI International, Inc.
• Original Assignee: Sarnoff Corporation (SRI International, Inc.)
• Inventors: Sajda, Paul, Pearson, John Carr, Spence, Clay Douglas
• Primary Examiner(s): Downs, Robert W.
• Application Number:
  US08/797,497
• Time in Patent Office: 1,082 Days
• Field of Search: 706/20, 706/25
• US Class Current: 706/20
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks
• CPC Class Codes: G06N 3/082 modifying the architecture, e.g. adding, deleting or silencing nodes or connections
• CPC Class Codes: G06T 7/0012 Biomedical image inspection
• CPC Class Codes: G06V 10/25 Determination of region of interest [ROI] or a volume of interest [VOI]
• CPC Class Codes: G06V 10/443 by matching or filtering
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 2201/03 Recognition of patterns in medical or anatomical images
• CPC Class Codes: G06V 30/194 References adjustable by an adaptive method, e.g. learning

Abstract

A signal processing apparatus and concomitant method for learning and integrating features from multiple resolutions for detecting and/or classifying objects are presented. Neural networks in a pattern tree structure with tree-structured descriptions of objects in terms of simple sub-patterns, are grown and trained to detect and integrate the sub-patterns. A plurality of objective functions and their approximations are presented to train the neural networks to detect sub-patterns of features of some class of objects. Objective functions for training neural networks to detect objects whose positions in the training data are uncertain and for addressing supervised learning where there are potential errors in the training data are also presented.

Claims

1. A method for growing a pattern tree having a root and at least one child, said method comprising the steps of:

(a) training the root of the pattern tree;

(b) training the children of the pattern tree; and

(c) creating at least one integration network, where said integration network receives its input from at least one of the children and root.

2. The method of claim 1, wherein said training step (a) applies ##EQU31## to train the root, where p_o\d is an expected value, y(x) is an output of a position x, N is a total number of pixels, n_o is a number of pixels inside of an object, u is an integration variable, and X_pos is a set of all positive positions.

3. The method of claim 1, wherein said training step (a) applies ##EQU32## to train the root, where E_DL is an error function, y(x) is an output of a position x, positives are positive regions and negatives are negative regions.

4. The method of claim 1, where said children and root are feature networks and wherein said creating step (c) creates an integration network for each of said feature network having children, where said integrated network receives inputs from said children of said integrated network'"'"'s mirror feature network and from said mirror feature network.

5. The method of claim 4, wherein said inputs to said integrated network is modified, where if a feature network has children which themselves have children, then said feature network'"'"'s mirror integration network is given input from said child feature networks'"'"' mirror integration networks, rather than from said feature networks themselves.

6. A pattern tree architecture of neural networks comprising:

a root feature network;

at least one child feature network coupled to said root feature network; and

at least one integration network, where said integration network receives its input from at least one of said children and root feature networks.

7. The pattern tree architecture of neural networks of claim 6, where each integration network is added for each of said feature network having children, where said integrated network receives inputs from said children of said integrated network'"'"'s mirror feature network and from said mirror feature network.

8. A method for training a neural network to discover features, said method comprising the step of:

(a) providing the neural network with a plurality of training data; and

(b) training the neural network using a function;

##EQU33## , where E_FD is an error function, where y(x) is an output of a position x, N is a total number of pixels, n_o is a number of pixels inside of an object, u is an integration variable and X_pos is a set of all positive positions.

9. The method of claim 8, wherein an approximation of said function is:

##EQU34## , where E_AFD is an approximation error function, where y(x) is an output of a position x, N is a total number of pixels, n_o is a number of pixels inside of an object, and X_pos is a set of all positive positions.

10. The method of claim 8, wherein a "blob-wise" approximation of said function is:

##EQU35## , where E_AFD is an approximation error function, N_b is a total number of blobs, n_pb is a number of positive blobs, ##EQU36## and y(x) is an output of a position x.

11. A method for training a neural network to detect objects with imprecise positions, said method comprising the step of:

(a) providing the neural network with a plurality of training data; and

(b) training the neural network using a function;

##EQU37## , where E_DL is an error function, y(x) is an output of a position x, positives are positive regions and negatives are negative regions.

12. A method for training a neural network, said method comprising the step of:

(a) providing the neural network with a plurality of the training data; and

(b) training the neural network using a function that accounts for errors in said training data, wherein said function is;

##EQU38##

13. A method for training a neural network, said method comprising the step of:

(a) providing the neural network with a plurality of the training data; and

(b) training the neural network using a function that accounts for errors in said training data, wherein said function is;

##EQU39##

14. A method for addressing a neural network trained with training data that contains error, said method comprising the step of:

(a) providing the neural network with a plurality of the training data that contains error; and

(b) correcting an output of the neural network using a corrected probability.

15. The method of claim 14 wherein said corrected probability is expressed as:

space="preserve" listing-type="equation">P(c=1|x)=π

.sub.1 y(x)+π

.sub.0 (1-y(x))

, where P(c=1|x) is said corrected probability, y(x) is an output for an i-th input x, π

₁ is a probability that said i-th input truly belongs to a first class 1 and π

₀ is a probability that said i-th input truly belongs to a second class 0.

16. The method of claim 14 wherein said corrected probability is expressed as:

##EQU40## , where P(C=c\x) is said corrected probability, y_d (x) is a network'"'"'s estimate of a probability of an input x being in classes N_c, and π

_c,d is a probability that said input truly belongs to said classes N_c.

17. A method for generating an integrated feature pyramid, said method comprising the steps of:

(a) generating a pyramid having a plurality of scales for each sample of an input signal;

(b) applying oriented filtering to each of said plurality of scales of said pyramid to produce a plurality of oriented output signals;

(c) squaring each of said plurality of oriented output signals to produce a squared output signal; and

(d) generating a pyramid having a plurality of scales for each of said squared output signal.

18. The method of claim 17, further comprising the step of:

sorting an energy of said oriented output signals by their magnitude at each pixel location prior to said generating step (d).

19. The method of claim 17, wherein said applying step (b) applies steerable filters.