Assignment 1 Reportx

ranchocucamongabrrrΤεχνίτη Νοημοσύνη και Ρομποτική

6 Νοε 2013 (πριν από 3 χρόνια και 5 μήνες)

61 εμφανίσεις



ENB441: Applied Image Processing

Case # 441
-
09

Assignment 1

Fingerprint and Suspect Photo
I
nvestigations

Team: JSquad

Name

Student ID

Ian Gilmour

05435102

Ben Wu

05434092


Ian for
Ben

Ben for Ian

Difference

5

5

0



Lecturer
:

Prof

Wageeh Boles

Due Date:
25 March 2008




-

i

-

Table of Contents


1

Executive Summary

................................
................................
................................
..................

1

2

Image Binarisation

................................
................................
................................
....................

2

2.1

Lite
rature Review

................................
................................
................................
.......................

2

2.1.1

Global Thresholding

................................
................................
................................
............

2

2.1.2

Adaptive Thresholding

................................
................................
................................
........

4

2.
2

Implementation and Discussion

................................
................................
................................
.

5

2.2.1

Suspect 1 Image Analysis

................................
................................
................................
....

5

2.2.2

Suspect 2 Image Analysis

................................
................................
................................
....

9

2.2.3

Suspect 3 Image Analysis

................................
................................
................................
..

11

2.2.4

Victim Image Analysis

................................
................................
................................
.......

14

3

Fingerprint Investigation

................................
................................
................................
........

16

3.1

Enhancement, edge detection and analysis
................................
................................
.............

16

3.1.1

Enhancement

................................
................................
................................
....................

16

3.1.2

Edge Detection

................................
................................
................................
..................

16

3.1.3

An
alysis

................................
................................
................................
..............................

19

3.1.4

Suspect Fingerprint 1

................................
................................
................................
........

20

3.1.5

Suspect Fingerprint 2

................................
................................
................................
........

21

3.1.6

Suspect Fingerprint 3

................................
................................
................................
........

21

3.2

Matching algorithm and analysis

................................
................................
.............................

22

3.2.1

Suspect Fingerprint 1

................................
................................
................................
........

23

3.2.2

Suspect Fingerprint 2

................................
................................
................................
........

25

3.2.3

Suspect Fingerprint 3

................................
................................
................................
........

27

3.2.4

Suspect Analysis Results

................................
................................
................................
...

29

4

Conclusions

................................
................................
................................
............................

31

4.1

Suspect Analysis

................................
................................
................................
.......................

31

4.2

Fing
erprint Analysis

................................
................................
................................
..................

31

5

References

................................
................................
................................
.............................

33


-

1

-

1

Executive Summary

The usefulness of image manipulation for forensic analysis is an invaluable asset to any investigator as
more often than not the evidence provided will be of poor quality and in need of cleaning up. This is
where the JSquad steps in and will be taking over the investigation of the case #441
-
09.


In this case of particular importance is the suspect images and
also fingerprint images, the report will
contain both a review on the methodologies of image binarisation and on fingerprint recognition
techniques used to identify the suspects at large.


For image binarisation
investigation into the use of types of thre
sholding and how they impact on
image restoration will be done in detail.


The fingerprint images received where of poor quality and the investigation into the use of both noise
reduction filters to improve the image quality and also Canny and Sobel metho
ds for edge detection
will be used to come to a conclusion.





-

2

-

Figure
2
: Image histogram. Distinct modes
(left), overlapping modes (center), indistingishable modes (right).

2

Image
B
inarisation

2.1

Literature Review

Image Binarisation is the process
of converting an
image that has a range of colours to a new image
that contains only two colour levels. The goal of
this process is to simplify the image to help
identify key elements within the original image. A
simple method for image binarisation is ima
ge
thresholding.

[1; 2; 3]



Thresholding is the process of selecting a colour level to separate the image upon. All pixels of a
colour less than the threshold are set to one colour (e.g. black) and all pixel
s of a colour greater than
or equal to the threshold are set to another colour (e.g. white). When a single threshold is used for an
entire image, the process is known as global thresholding. If multiple thresholds are used throughout
the image the process
is called adaptive or local thresholding.
[1; 2]

2.1.1

Global Thresholding

To effectively select a global threshold value, there needs to be two ‘distinct modes’ within the image.
A distinct mode must be made up of a range of colours that are limited to one part of the colour
spectrum and that do not overlap with the range of col
ours in the other mode.

There are many methods for selecting a global threshold. For simple images, a trial and error

technique can be implemented or, for a more directed approach, the image histogram can be
inspected and thresholds that match histogram minima can be trialled (as this is the most probable
area for overlapping modes). Automatically identifying the best mi
nima to be used as the threshold is
the focus of many image processing algorithms; two such methods are ‘Otsu’s method’ and an
iterative procedure described in “Digital image processing using MATLAB”

[1;
2; 3; 4]
.

Figure
1
: Image binarisation (histogram before and after)
[2]


-

3

-

Figure
3
: Original (left); Otsu Thresholding (right)

[3]

Otsu’s method works by assuming that an image is bimodal (two histogram modes) and that if intra
-
class variance of each mode i
s minimised then the threshold will be most effective. Intra
-
class
variance is the variance of each mode independent of the other


i.e. class A variance is minimised
with respect to mean A and class B variance is minimised with
respect to mean B. This i
s achieved by stepping through all
possible thresholds and testing if the intra
-
class variance is
being minimised

[4]
.


Another method that works on a similar principle as Otsu’s method is an iterative method described
in “Dig
ital image processing using MATLAB.” The method requires less computational power and in a
lot of cases will return a result that is as good as Otsu’s method. However there are some cases where
a better threshold exists but will not be found by the algorit
hm

[2; 3]
. The algorithm requires
repeated mean calculations as follows:



Set Threshold as the image mean



Loop until threshold does not change in successive iterations

o

Calculate mode means using current threshold

o

Set th
reshold as average of means


Global thresholding is simple and effective for some images; however, when an image does not have
distinct colour modes it is generally ineffective. An image that has been exposed to uneven
illumination is an example when globa
l thresholding will probably fail (for example
Figure
4
). This
failing can be overcome by using adaptive or local thresholding

[5;
2]
.




Figure
4
: Uneven Illumination (left)
; Global Thresholding (right)

[5]


-

4

-

2.1.2

Adaptive Thresholding

Adaptive thresholding simply indicates that the threshold used to binarise the image is not
constant
across all parts of the image. Local thresholding is a subclass of adaptive thresholding and indicates
that thresholds are calculated for each pixel based on the local surrounding pixels. Local thresholding
works much the same as global threshold
ing except that it is based on the assumption that a small
area of an image will have relatively constant illumination and hence will overcome the problem faced
in global thresholding

[5; 4]
.


In local thresholding automatic selection of a threshold is required and
the same algorithms that were used for global thresholding can be used.
For example, when the Otsu’s method is used to select the threshold in
a sequential system, a kernel is slid ac
ross the image and the threshold
for the centre pixel is chosen based on the values within the kernel. This
is a very effective type of local thresholding but, on a sequential
machine, it is also extremely computationally expensive.


Alternatively, stati
stical methods that are not as computationally heavy can be used. The same
process as described for the adaptive Otsu method is used except that the threshold is chosen by
taking a statistical function of the kernel, for example the mean or the medium

[5]
.


When using adaptive thresholding, window (kernel) size will affect the performance and cost of the
threshold. A small kernel will be fast and pick up the most detail but will also be more susceptible to
noise and if too sma
ll may not cover both modes of the image. A large kernel will be slow and less
susceptible to noise but will be closer to having the same problems as global thresholding and, also,
will remove small details

[4]
.







Figure
5
: Adaptive Thresholding
[5]

Figure

6
: Original (left); Global Thresholding (centre); Adaptive Thresholding (right)
[2]


-

5

-

2.2

Implementation and Discussion

2.2.1

Suspect 1 Image Analysis



Original Suspect Image

Histogram Equalised Suspect
Image

To make the image clearer to the human eye and to allow for easier identifications of histogram
minima, the original image was histogram equalised. From this simple operation it can be see that the
suspect is wearing a shirt that has a ‘QUT’ logo. I
n an attempt to obtain the suspects
silhouette, the
image was binar
ised using global thresholding. Note the hair cut and what seems to be glasses.



Equalised Image histogram with manual
threshold

Image with threshold 100 applied.


-

6

-



Equalised Image histogram with iterative threshold

Image with threshold 122 applied.



Equalised Image histogram with
Otsu

threshold

Image with threshold 121 applied.

The image has fairly consistent illumination, so the silhouettes achieved from global thresholding
should not see a signification improvement from adaptive thresholding.


-

7

-




Adaptive Thresholding 7x7

Adaptive Thresholding 39x39

Adaptive Thresholding 71x71




Adaptive Thresholding 27x27

Adaptive Thresholding 67x67

Adaptive Thresholding 107x107

When adaptive thresholding is applied to the suspect 1 image, no new evidence is found. The
adaptive threshold using a mean filter of size 71x71 does give a clean silhouette of the suspect’s head.
The smaller window sizes allow more detail to be seen howev
er this detail is often confusing and
distracting from the important parts of the image.


-

8

-




Otsu Adaptive Thresholding

When Otsu adaptive thresholding is applied to the image of suspect 1, a detailed view of the scene is
shown however this comes at great processing cost (each thresholding procedure took over 3
minutes).


When global and adaptive thresholding is used to an
alysis the image of suspe
ct 1, an accurate
silhouette of the suspect is obtained. The global thresholds yield a filled silhouette showing the
suspects form whilst the adaptive thresholding processes yield a detailed silhouette that reveals
evidence such as

clothing form, hair cut and some facial features. It was also noted that the suspect
was wearing a shirt with an inscribed logo reading ‘QUT’.




-

9

-

2.2.2

Suspect 2 Image Analysis



Original Suspect Image

Histogram Equalised
Suspect Image

From the histogram equalisation it can be see that the suspect is wearing a shirt that has an ‘IPT’ logo.



Equalised Image histogram with manual threshold

Image with threshold
207

applied.



Equalised Image histogram with iterative threshold

Image with threshold 12
9

applied.


-

10

-



Equalised Image histogram with
Otsu

threshold

Image with threshold
88

applied.

Global thresholding is ineffective on the image of suspect 2 as the suspect has uneven illumination
from the light sources at the scene.
No global threshold allows a clear silhouette of the suspect and so
adaptive thresholding must be used to gather furthe
r evidence.




Local thresholding using Mean criteria




Local thresholding using Medium criteria

The mean and medium statistical method for local thresholding only produces rough silhouettes of
the suspect. The rapid change in lighting across the face of the suspect makes it difficult for
thresholding to be used effectively.





-

11

-




Local thresholding using Otsu’s method

周攠Ot獵 aT慰t楶攠tU牥sU潬oin朠浥tU潤 祩e汤猠tU攠b敳t 牥su汴s fo爠tUe 業i来g潦⁳o獰散e ㈮2䥴 a汬lw猠愠
rough silhouette of the suspects face to be obtained as well as a rough view of the suspect’s hair style.
In addition to this, a clear logo ‘IPT’ can be s
een written on the back side of the suspect’s shirt.

2.2.3

Suspect 3 Image Analysis



Original Suspect Image

Histogram Equalised Suspect Image




Equalised Image histogram with manual
threshold

Image with threshold

25

applied.


-

12

-



Equalised Image histogram with iterative threshold

Image with threshold 1
03
applied.



Equalised Image histogram with
Otsu

threshold

Image with threshold
91

applied.





Local thresholding using Mean criteria


-

13

-




Local thresholding using Medium criteria




Local thresholding using Otsu’s method

周攠i浡来 of su獰散e 3 楳i愠Tiff楣u汴 業慧i t漠p牯c敳猠u獩sg 業慧i tUr敳e潬T楮朮⁂潴o 杬gb慬a慮T
慤慰t楶攠tU牥sU潬T楮朠祩敬e T楳慰p潩湴楮g 牥獵lt献 周楳i楳 Tu攠t漠tU攠f潲敧e潵nT
⡳u獰散e b潤礩 b敩n朠
tUe 獡浥 c潬潵r 慳atUe

浡m潲楴礠潦 tU攠b慣a杲潵gT⸠
TU牥獨潬oin朠牥qui牥猠tUat tUe牥 楳 愠T楦f敲enc攠楮
c潬ou爠b整w敥n T楦f敲ent 浯T敳 潦⁴U攠業慧i anT tU楳 w慳a污捫楮朠楮 tU攠i浡来 of 獵sp散e ㌮3
䄠牯u杨
獩汨潵ett攠of su獰散e ㌠楳 潢瑡i
neT but it 楳inot c汥慲a潲 T散楳楶攮eN漠otU敲 敶楤敮ce c潵lT b攠t慫敮
f牯洠tU楳 業慧i⸠


-

14

-

2.2.4

Victim Image Analysis



Original Victim Image

Histogram Equalised Victim Image

The image of the victim was analysed using thresholding to check for hidden evidence.



Equalised Image histogram with
Otsu

threshold

Image with threshold 1
03

applied.



-

15

-










Both thresholding techniques did not yield any new evidence. The only
result of the analysis was a
good silhouette of the victim but this is not helpful for the case.



-

16

-

3

Fingerprint
Investigation

3.1

Enhancement, edge
detection and analysis

3.1.1

Enhancement

For initial image enhancement due to all of the fingerprints being affected heavily by noise a median
filter is used to clean up the prints. This cleans up the images for the following edge detection.




As you can see

although the image is blurred slightly after applying a median filter it greater reduces
the noise and such that when used with our fingerprints detection allows for a dramatic increase in
fingerprint detection accuracy. From the image above you can see t
hat the median filter works by
replacing the pixel in question with the median value of its surrounding
pixels.
[6]


Edge detection is used in image processing particularly for detecting and extracting features within
the image
. Various algorithms have been devised which
attempt to solve this problem of these
algorithms they can be primarily be split into two major types; gradient based and derivative based.


3.1.2

Edge Detection

For the following analysis we will be using both the S
obel and Canny methods in order to extract
these edges from the various fingerprints supplied and also implement a matching algorithm to
determine the suspect. These methods will provide the best solution as they cover the two major
edge detection algorith
ms used.
[7]



-

17

-

For Sobel edge

detection the primary method put in place is a gradient detection algorithm which
detects edges by taking the first derivative of the input and should the gradient value of this now
exceed a

predetermined threshold an edge is detected. The below images represent this method with
a simple one dimensional signal.



As you can see once the original signal (Right) is derived the rate of change should this surpass your
threshold defines an edge.
This once represented in 2
-
d results now uses a 3x3 convolution mask to
estimate which passes through pixels manipulating each

square at a time changing their values.
[8]


For canny edge detection there are multiple steps the first involves using a Gaussian filter which
reduces the noise and smooths out the image using the following formula.
[9]


Where


The next stage is to use any of the gradient operators (Roberts, Sobel, Prewitt) to identify the edge
strengths. The Sobel mask shown above is a prime example of what can be used at this stage. Once
this is done the
angles of the edges can now be identified.


-

18

-


Edges can be defined as above using colours to determine at what angle the pixels in the related
neighbourhood connect.


Original Image


Canny Image using different thresholds




-

19

-

An example of how both the So
bel and Canny methods result in is shown below.
[9]



Original





Sobel Method




Canny Method

3.1.3

Analysis

From inspecting the fingerprint database it is apparent that the fingerprints align up and only
distorted by noise. Thu
s once the images for both the suspects and the candidates are median filtered
and passed through with the Canny and Sobel algorithms it is possible just to compare the images in
Matlab by using the ‘and’ function which results in showing us which parts of

the image do not align.

It is through this approach that we can identify the suspect with the highest percentage match.




-

20

-

3.1.4

Suspect Fingerprint 1

For each of the suspects the original fingerprint is smoothed with the median filter than the Canny
and Sobel

algorithms are run on the result.

Canny Result:


Sobel Result:


Results show that the Canny filter removes most of the noise however at the cost of some detail, and
the Sobel method includes all the detail however also not filtering out the noise. Different thresholds
where attempted and these are the best results.



-

21

-

3.1.5

S
uspect Fingerprint 2

Canny result:



Sobel result:


Again similar results as the first suspect’s image with the Sobel resulting in more noise although
maximum detail.


3.1.6

Suspect Fingerprint 3

Canny Result:


-

22

-


Sobel Result:


The noise in the Sobel result
is extensive and results in a very confusing result however noise aside
the fingerprint lines are still present allowing this result to be used.

3.2

Matching algorithm and analysis

For the analysis once the fingerprint is filtered and edge detected using both

Canny and Sobel
methods the results are joined together with only the matching pixels remaining. The suspect with
the highest % is then chosen as the primary suspect; however for complete satisfaction that the result
is genuine the next highset candidate
is also show to prove the result.




-

23

-

3.2.1

Suspect Fingerprint 1

Canny:



Result:

Person 40 matches fingerprint 1

Person 39 is the next closest candidate


-

24

-

Sobel:



Result:

Person 40 matches fingerprint 1

Person 39 is the next closest candidate



-

25

-

3.2.2

Suspect
Fingerprint 2

Canny Result:



Result:

Person 15 is primary suspect

Person 36 is secondary suspect.


-

26

-

Sobel Result:



Result:

Person 15 is primary suspect

Person 36 is secondary suspect.


-

27

-

3.2.3

Suspect Fingerprint 3

Canny:



Result:

Primary suspect is person

40

Secondary suspect is person 39



-

28

-

Sobel:



Result:

Primary suspect is person 40

Secondary suspect is person 39




-

29

-

3.2.4

Suspect Analysis Results


Canny Method Results:

Fingerprint:

Primary Suspect

%

Secondary Suspect

%

1

Person 40

55.7

Person 39

21.6

2

Person 15

60.3

Person 36

20.8

3

Person 39

32.9

Person 36

20.1



Sobel Method Results:

Fingerprint:

Primary Suspect

%

Secondary Suspect

%

1

Person 40

38.3

Person 39

24.2

2

Person 15

42.7

Person 36

24.7

3

Person 39

27.0

Person 36

22.6





-

30

-

From the
results it proves conclusively that the suspects are as follows:

Fingerprint 1

Person 40



Fingerprint 2

Person 15



Fingerprint 3

Person 39




-

31

-

4

Conclusions

4.1

Suspect Analysis

The suspect analysis of the images proved to us that through the many methods and image analysis
that we can extract data which is originally hidden from view. From the results of the analysis it can
be seen that from images 1 and 2 that they are different

people however image 3 does not yield too
much extra information about the suspect.

The use of thresholding was a key technique used throughout the process with it giving us the ability
to differentiate background from foreground objects with ease.
The v
arious types of thresholding are
useful in different situations with each type giving us improved results.


4.2

Fingerprint Analysis

For the fingerprint recognition application it is visible that due to the excessive noise in the fingerprint
files that it is better to use the Canny method over the Sobel as the Canny will accept a much wider
threshold values allowing for more or less dat
a. The Sobel method although still successfully
identifying the suspect’s results in a very noisy filtered image this in turn made the percentage gap
between the closest suspects a lot closer.

This comes down to how the methods themselves operate and the
Canny algorithm process is much
longer however provides a clearer edge in the final image, the Sobel however maintains large amount
of the original image when detecting edges and can be problematic when there is excessive noise like
in this case.





-

32

-

The a
ctual suspect at this time cannot definitely be deduced given the images however upon piecing
together the clues I am hesitant to say that this is an inside job to stop the victim from speaking out to
us about the atrocities that must have been occurring.
The QUT logo and the IPT logo which I can
assume mean Image Processing Technologies all point towards a certain teaching staff within the
Image Processing team. I trust that you treat this material sensitively as should sound of this leak out
we may fright
en the attacker into hiding. With that said I am tentatively going to say that the suspect
to which we should investigate further points towards one Brenden Chen given his involvement with
the Image processing team as he is the only one who has the skill t
o manipulate the images supplied
so we cannot draw conclusive evidence yet still maintain the facade of attempting to solve the case.
The hairstyle, glasses and face shape all point further towards Chen. I highly doubt any original image
could have been ma
lformed so dramatically without some outside intervention my only hope is that
this information reaches the right person before we have been found out.






-

33

-

5

References

[1] Fisher, Robert, et al., Thresholding.
Hypermedia Image Processing Ref
erence.
[Online] 2003.
http://homepages.inf.ed.ac.uk/rbf/HIPR2/threshld.htm.

[2] Gonzalez, Rafael C. and Woods, Richard E.,
Digital Image Processing.
2nd. New Jersey

: Prentice
Hall, 2002.

[3] Gonzalez, Rafael C., Woods, Richard E. and Eddins, Steven L.,
Digital image processing using
MATLAB.
New Jersey

: Prentice Hall, 2004.

[4] Morse, B., "Thresholding." [Online] 2000.
http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/MORSE/threshold.pdf.

[5] Fisher, Robert, et al., Adaptive Thresholding.
Hypermedi
a Image Processing Reference.
[Online]
2003. http://homepages.inf.ed.ac.uk/rbf/HIPR2/adpthrsh.htm.

[6] R. Fisher, S. Perkins, A. Walker, E. Wolfart., Median Filter. [Online] 2003. [Cited: 25 4 2009.]
http://homepages.inf.ed.ac.uk/rbf/HIPR2/median.htm.

[7]
Rowan, Jim., Edge detection. [Online] 1998. [Cited: 25 4 2009.]
http://www.cc.gatech.edu/classes/cs7321_98_winter/participants/jrowan/ps2/.

[8] Green, Bill., Edge Detection Tutorial.
Edge Detection Tutorial.
[Online] 2002. [Cited: 25 4 2009.]
http://www.pa
ges.drexel.edu/~weg22/edge.html.

[9] Ruye, Wang., Canny Edge Detection. [Online] 2003. [Cited: 25 4 2009.]
http://fourier.eng.hmc.edu/e161/lectures/canny/node1.html.

[10] Green, Bill., Canny Edge Detection Tutorial. [Online] 2002. [Cited: 25 4 2009.]
http:
//www.pages.drexel.edu/~weg22/can_tut.html.