14
A Correlation Based Approach to Quality and Noise in Crime Scene Fingerprint Duy Hoang Thai, Lucas Mentch and Zhengwu Zhang 1

A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

Page 1: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

A Correlation Based Approach to Quality andNoise in Crime Scene Fingerprint

Duy Hoang Thai, Lucas Mentch and Zhengwu Zhang

1

Page 2: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

Overview:

1. Motivation of a quality metric and noise modeling forfingerprinting

2. Analysis for quality metrics and noise modeling

3. Computational tool

4. Question and hints

2

Page 3: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

1. Motivation of a quality metric and noise modeling for fingerprinting

Example: FVC 2000 database consists of three databases which arecaptured from three different sensors:

I DB1: low-cost optical sensor ”Secure Desktop Scanner” byKeyTronic

I DB2: low-cost capacitive sensor ”TouchChip” by STMicroelectronics

I DB2: optical sensor ”DF-90” by Identicator Technology

3

Page 4: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

1. Motivation of a quality metric and noise modeling for fingerprinting

Observation:

I Every image is captured from different sensors.

I Every sensor produces different quality image and noise.

I There are smudge and/or dryness on fingerprint

(a) dryness (b) wetness

Thus, we need to find quality metrics to evaluate the image quality offingerprinting and statistically model noise produced by different sensors.

4

Page 5: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

2. Analysis for quality metrics and noise modeling

good texture

bad area(smudge)

"correlated" noise

There are three features to define quality metrics:

1. areas with good texture, i.e. the green,

2. areas with bad texture but still in the region of interest, i.e. the red,

3. small scale (correlated) noise outside the ROI which is caused bydifferent kind of sensors.

5

Page 6: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

2. Analysis a quality metric and noise modeling

Questions:

1. How can we define the region of interest (ROI) anddistinguish the good areas (green) and the bad areas(red)?

2. How can we define quality metrics to evaluate thecorrelation between the good and bad areas?

3. How can we statistically model the ”correlated” noiseoutside the ROI?

6

Page 7: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

2. Analysis a quality metric and noise modeling

For question 1:

I We can apply the idea from the decomposition method todistinguish the textural and non-textural areas on the fingerprintimage which is based on [1, 2].

I Then, we easily define the ROI, the good areas and the bad areason the fingerprint images.

Note:

I There are several approaches to solve this problem anddecompostion [1, 2] is only one of these tools.

I The source code for this step is provided.

—————————————————————1 D. H. Thai and C. Gottschlich, Global variational method for fingerprint segmentation by three-part

decomposition, IET Biometrics 2 (5), 120 – 130, 2016.

2 —————————————, Directional global three-part image decomposition, EURASIP Journal onImage and Video Processing 12 (2016), 1 – 20.

7

Page 8: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

3. Computational tool

Computational tool: Decomposition for fingerprint image [1, 2]

Mathematical model:

min(u,v,ε)∈X 3

{∥∥∥∇+L u∥∥∥`1

+ µ1

∥∥C{v}∥∥GS

+ µ2‖v‖`1s.t.

∥∥C{ε}∥∥`∞≤ δ , f = u + v + ε .

}Note: parameters µ1, µ2 and δ are trained in [1]. There are some fingerprint patternsin u. It’s fine because we only care binarized texture v and residual (noise) ε.

8

Page 9: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

3. Computational tool: Matlab software [1]

demo.m, Segmentation G3PD.m

clcclear all

% Read image:name = 'FVC2000_DB3_IM_4_1.png'; f = double( imread(name) ); f = f(:,:,1);

% Parameters:mu1 = 1; beta1 = 1e-3; beta3 = 1e-3; gamma = 1e-3; dim = 9; const = 5; NumBlock = 6; padsize = 15;mu2 = 0.055; beta2 = 0.001; Iteration = 4; % FVC2000_DB3

% Segmentation: [ROI, Mask, v, u, n, Image_seg, Error_v] = Segmentation_G3PD(f, mu1, mu2, Iteration, beta1, beta2, beta3, gamma, dim, const, NumBlock, padsize);Mask = ~Mask;

f_draw1 = DrawBoundary(f, ROI);f_draw2 = DrawBoundary1(f_draw1, Mask);

figure, imshow(uint8(f_draw2))

Note:

I image f is a matrix whose value of every pixel is a real number R.

I The value of parameters for different FVC2000 databases areprovided in [1, p 15]: http://arxiv.org/pdf/1505.04585.pdf.

9

Page 10: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

3. Computational tool

Numerical result for demo.m [3]:By the decomposition method [1, 2], we can define the ROI, thegood areas and the bad areas as

(a) segmented image (b) ROI (c) Smudge

—————————————————————3 L. Mentch et. al., An Automated Multiscore Approach to Assessing the Quality of Latent Pattern Evidence

(ongoing).

10

Page 11: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

3. Computational tool

Then, one possible approach to define a quality metric is amethod of local and smooth variance for the contrastmeasurement:

(a) Variance contrast

0 1000 2000 3000 4000 5000 6000 7000 8000 9000

#10 -4

0

0.5

1

1.5

2

2.5

3

3.5

(b) Snoothed histogram of(a) (the red curve)

11

Page 12: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

4. Question and hints

Questions:Given a fingerprint image

how do we define the quality metric and find a statisticalmodel of correlated noise?

Database FVC2000: DB1, DB2 and DB3http://bias.csr.unibo.it/fvc2000/download.asp

Recommended software: demo.m, Segmentation G3PD.m

12

Page 13: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

4. Question and hints

Hints:

1. The quality metric: compute the contrast measure by meanor variance of the texture of fingerprint pattern over theregion of interest in a fingerprint image.

2. Statistical model of correlated noise: model noise bystochastic process, e.g. ARMA process, and compute itsstatistical properties.

13

Page 14: A Correlation Based Approach to Quality and Noise in Crime ... · I Every image is captured from di erent sensors. I Every sensor produces di erent quality image and noise. I There

Thank you!

14