S. Rahnamayan, H.R. Tizhoosh, M.M.A. Salama University of Waterloo, Ontario, Canada

S. Rahnamayan, H.R. Tizhoosh, M.M.A. Salama

University of Waterloo, Ontario, Canada

MOPTA, Windsor, July 26, 2005

Optimization of Object Extraction Based on One User-Prepared Sample

2

Agenda

Objective Proposed Approach Preliminary Results Comparison with Other Methods Conclusion Future Works

3

Main Objective

Acquisition of object extraction procedure from user-prepared sample(s) based on genetic

optimization of morphological processing chains

4

Reasons for Developing Automated Image Processing Systems

Dealing with huge number of images Saving experts valuable time Possibility of using in online

applications Overcoming of inconsistent nature of

human processing Supporting required high accuracy

5

Why Learning from a Small Number of Samples is Valuable ?

Because :

It reduces the expected level of expert participation which is the main obstacle for research and development. Preparing some manually generated samples to reflect the experts’ expectations is a reasonable requirement in all image processing environments.

6

Proposed Approach

Utilizing Mathematical morphology operations, as image processing tools, to build object extraction procedure

Using genetic algorithm, as optimizer tools, to find optimal parameters of above mentioned procedure

7

Morphological Operations

They are shape-based operations

Used to handle a wide range of image processing tasks, ranging from noise filtering to object extraction

8

Procedure ApplierProcedure Applier

InputImage Optimal

Parameters

InputImages

Resultimages

Gold Image

Optimal Ordering ofOperations

Genetic OptimizerGenetic Optimizer

Parameter & OrderingOptimizer

MathematicalMorphologyOperations

Main Structure of Proposed Approach

9

Morphological Operations Chain as a Morphological Procedure

SE1, SE2, SE3, and SE4: Corresponding structural elementsK1, K2, and K3 : Iteration times for operationsO: OpeningC: ClosingD: DilationE: Erosion

1. K3*{O(SE1)_C(SE2)}K1*E(SE3)K2*D(SE4)

2. K1*E(SE3)K3*{O(SE1)_C(SE2)}K2*D(SE4)

3. K1*E(SE3)K2*D(SE4)K3*{O(SE1)_C(SE2)}

4. K3*{O(SE1)_C(SE2)}K2*D(SE4)K1*E(SE3)

5. K3*{O(SE1)_C(SE2)}K2*D(SE4)K1*E(SE3)

6. K3*{O(SE1)_C(SE2)}K2*D(SE4)K1*D(SE3)

10

Genetic Optimization of MM Procedure

PopulationInitialization

ApplyingMM Procedure

Computing ofDissimilarity

Selection

Is Reached Ending Criteria?

Crossover

Mutation

Start

EndYes

No

11

Preliminary Results

Circle Extraction Triangle Extraction Rectangle Extraction Object Extraction Applied for Grey-level

Images

12

Utilized Measures

Matching Index:

Overall Matching Index:

13

(a) Original image(b) Goal image(c) Generated image by MM procedure (94.48% similarity)

Training for Object Extraction- Circle

14

Improvement of Result Performance During Training

15

Object (Circle) Extraction Training Results

16

Verification of Optimization

17

Results of Object (Circle) Extraction

18

(a) Original image(b) Goal image(c) Generated image by MM procedure (85.01% similarity)

Training for Object Extraction- Triangle

19

Results of Object (Triangle) Extraction

20

(a) Original image(b) Goal image(c) Generated image by MM procedure (94.37% similarity)

Training for Object Extraction- Rectangle

21

Results of Object (Rectangle) Extraction

22

Summary of Numerical Results

23

Object Extraction Applied on Gray-scale Images

(a) Grey scale image(b) Goal image(c) Generated image by MM procedure (76.77% similarity)

24

Some Results of Object Extraction in Grey Level Images

95.05% 96.05%

96.09% 96.71%

95.63%

Overall matching rate: 95.90% with standard deviation of 0.54%

25

Level of supported variations

Noise Adding

Translating Duplicating

Overlapping

Scaling Rotating

Partial Complete Complete High Partial Partial

26

Training for Fully Rotation Invariant Triangle Extraction

GeneticOptimizerGenetic

Optimizer

1

2

3

4

1 2 3 4

Inp

ut Im

ag

es

1

2

3

4

Resu

lt Imag

es

27

Comparing Proposed Approach with Knowledge-Based Learning

Knowledge acquisition difficulties √ Unable of self-learning √ Difficult to avoid conflicts in large knowledge bases √ Knowledge reliability problem √

√ : Proposed approach solves it mostly or it is not applicable.

28

Sample providing problem √ Problem of choosing the best architecture √

Comparing Proposed Approach with Sample-Based (NN) Learning

~ : Proposed approach solves it partially.

29

Conclusion

The outstanding features of the proposed approach are as follows:

- Training based on a few samples - Supporting (semi) automated image processing- Mostly invariant for noising, overlapping,

translation, rotating, scaling, and duplicating.

30

Future Works

- Extending functionality of the system to cover wider range of image processing tasks

- Applying on medical image processing

31

Thank you for your attention and patience.