Methods of improving detectors and classifiers using optimized stochastic resonance noise

• US 9,026,404 B2
• Filed: 03/02/2012
• Issued: 05/05/2015
• Est. Priority Date: 10/20/2005
• Status: Active Grant

• Current Assignee: Syracuse University
• Original Assignee: Syracuse University
• Inventors: Peng, Renbin, Varshney, Pramod K.
• Primary Examiner(s): Huynh, Phuong
• Application Number:
  US13/410,949
• Publication Number:
  US 20120278039A1
• Time in Patent Office: 1,159 Days
• Field of Search: 702179-181, 702/191
• US Class Current: 702/191
• CPC Class Codes: A61B 6/502 for diagnosis of breast, i.e. mammography
• CPC Class Codes: G06T 2207/10116 X-ray image
• CPC Class Codes: G06T 2207/30068 Mammography; Breast
• CPC Class Codes: G06T 5/20 using local operators
• CPC Class Codes: G06T 5/70 Denoising; Smoothing
• CPC Class Codes: H04B 17/26 using historical data, averaging values or statistics

Abstract

Apparatus and method for improving the performance of a threshold-based detector or classifier, or a generic detector or classifier and increasing the probability of detecting at least one object in an image using novel algorithms and stochastic resonance noise is provided, where a suitable dose of noise is introduced to the image data such that the performance of the above-referenced detectors or classifiers is improved without altering the detector'"'"'s or classifier'"'"'s parameters. Several stochastic resonance (SR) noise-based detection and classification enhancement schemes are presented. The SR noise-enhanced detection and classification schemes can improve any algorithms and systems. To implement these schemes, the only knowledge that is needed is the original input data (no matter 1D, 2D, 3D or others) and the output (detection results) of the existing algorithms and systems.

Claims

1. A method of improving the performance of a threshold-based detector or classifier, and increasing the probability of detecting or classifying at least one object in an image, said method comprising the steps of:

a. obtaining image data by using an imaging system, wherein said image data comprises a first set of positive pixels and negative pixels;

b. calculating stochastic resonance noise with a processor by determining the stochastic resonance noise probability density function that maximizes the probability of detection while not increasing the probability of false alarm, and calculating the stochastic resonance noise data probability density function from a known probability density function for said image data, wherein the stochastic resonance noise probability density function can have the form of p_n^opt(n)=λ

δ

(n−

n₁)+(1−

λ

)δ

(n−

n₂), with values n₁ and n₂ equal to two delta function locations having probabilities of λ and

1−

λ

respectively, where n₁ and n₂ denote two discrete vectors, λ

denotes the probability of occurrence of said two vectors, and n is a randomization of said two discrete vectors added with the probabilities of λ and

1−

λ

, respectively; and

c. adding said stochastic resonance noise to said image data by using said processor, thereby improving the performance of a threshold-based detector or classifier and increasing the probability of detecting or classifying at least one object in an image.

2. The method of claim 1, further comprising the step of calculating a detection or classification threshold for said threshold-based detector or classifier.

3. The method of claim 2, further comprising obtaining a first set of probability density estimates using the detected positive pixels and negative pixels.

4. The method of claim 3, wherein the step of obtaining said first set of probability density estimates using the detected first set of positive pixels and negative pixels further comprises the step of obtaining said first set of probability density estimates under the following two hypotheses:

5. The method of claim 4, further comprising the step of updating said detection or classification threshold for said threshold-based detector or classifier according to said first set of probability density estimates.

6. The method of claim 5, further comprising the step of detecting a second set of positive pixels and negative pixels with said updated detection threshold.

7. The method of claim 6, further comprising the step of obtaining a second set of probability density estimates using the detected second set of positive pixels and negative pixels.

8. The method of claim 3, wherein the step of obtaining said second set of probability density estimates using the detected second set of positive pixels and negative pixels further comprises the step of obtaining said second set of probability density estimates under the following two hypotheses:

9. The method of claim 8, wherein the step of calculating stochastic resonance noise further comprises the step of using said updated detection or classification threshold and said second set of probability density estimates to calculate the stochastic resonance noise.

10. The method of claim 9, further comprising the step of performing the step of detection or classification on the stochastic resonance noise modified image data.

11. The method of claim 10, wherein each of said steps are performed in an iterative manner.

12. The method of claim 10, wherein said steps are repeated in an iterative manner if the difference between two successive detection or classification results is larger than a pre-set threshold value.

13. The method of claim 4, further comprising the step of performing a test to choose between the two hypotheses, which is characterized by a critical function, or decision function, φ

( y), 0≦

φ

( y)≦

1, and

14. The method of claim 13, wherein a detection or classification performance of said test is evaluated in terms of probability of detection (P_D) and probability of false alarm (P_F),

15. The method of claim 14, wherein said stochastic resonance noise is represented by n, and is added to the original data data y, which yields a new data vector z as follows

z= y+ n

where n is either a random vector with probability density function p_n(.) or a non-random signal.

16. A method of improving the performance of a generic detector or classifier, and increasing the probability of detecting or classifying at least one object in an image, said method comprising the steps of:

a. obtaining image data by using an imaging system;

b. calculating optimum stochastic resonance noise with a processor, wherein said optimum stochastic resonance noise comprises two-peak stochastic resonance noise having the following expression p_n^opt( n)=λ

δ

( n−

n₁)+(1−

λ

)δ

( n−

n₂), where λ and

1−

λ

are occurrence probabilities of the suitable N-dimensional vectors n₁ and n₂, and 0≦

λ

≦

1, and where said optimum stochastic resonance noise includes three parameters, n₁, n₂ and λ

; and

c. adding said optimum stochastic resonance noise to said image data by using said processor, thereby improving the performance of a detector and classifier and increasing the probability of detecting or classifying at least one object in an image.

17. The method of claim 16, wherein said optimum stochastic resonance noise is represented by n, and is added to the original data data y, which yields a new data vector z as follows

z= y+ n

where n is either a random vector with probability density function p_n(.) or a non-random signal.

18. The method of claim 17, further comprising the step of maximizing P_D^z under the constraint that P_F^z≦

P_F^y, where y and z denote the original and stochastic resonance noise-modified data, respectively, and where P_D denotes the probability of detection and P_F denotes probability of a false positive.

19. The method of claim 16, wherein said image data comprises a first set of negative or background pixels, and further comprising the step of:

a. estimating a probability density function, p₀, of background pixels in the original image using said obtained first set of negative or background pixels under the following hypothesis;

H₀;

p_y(y;

H₀)=p₀(y)

where y is an N-dimensional data vector, yε

R^N, p₀( y) is a probability density function of y under H₀ representing a negative pixel hypothesis.

20. The method of claim 19, further comprising the step of maximizing a number of detected positive pixels under the constraint that P_F^z≦

P_F^y, where y and z denote the original and stochastic resonance noise-modified data, respectively, and where P_D denotes the probability of detection and P_F denotes probability of a false positive.

21. The method of claim 16, further comprising the step of detecting fusion by repeating a performance of the step of detection on the stochastic resonance noise modified image data K times with the same stochastic resonance noise parameters corresponding to K different realizations of said stochastic resonance noise-modified images;

and the step of recording the number of times a pixel is detected in said stochastic resonance noise modified image data correlating to said object during said K times, wherein a number of detected positive pixels is equal to the number of pixels whose recorded times of correlating to said object is larger than or equal to Q, where K and Q are positive integers, and Q≦

K .

22. The method of claim 16, wherein said image data comprises a first set of negative or background pixels, and further comprising the step of:

a. estimating a probability density function, p₀, of background pixels in the original image from the detected negative or background pixels, denoted by p₀^est, which is fused with a prior probability density function, denoted by p₀^prior, using the following convex combination;

23. The method of claim 22, further comprising the step of detecting fusion by repeating a performance of the step of detection on the stochastic resonance noise modified image data K times with the same stochastic resonance noise parameters corresponding to K different realizations of said stochastic resonance noise-modified images;

and the step of recording the number of times a pixel is detected in said stochastic resonance noise modified image data correlating to said anomaly during said K times, wherein a number of detected positive pixels is equal to the number of pixels whose recorded times of correlating to said anomaly is larger than or equal to Q, where K and Q are positive integers, and Q≦

K.