Embed Size (px)
Citation preview
Hidden Markov Random Fields and Swarm Particles: a Winning Combination in Image Segmentation
2014 International Conference on Future
Information Engineering
EL-Hachemi Guerrout,Teacher at ESI
Samy Ait-Aoudia, Professor at ESI
Ramdane Mahiou, Teacher at ESI
The goal of segmentation:
To simplify the representation of an image
The image became more meaningful and easier to analyze
Introduction- Why the segmentation ?
2
Medical imagingLocate tumors Diagnosis pathologies…
Object detection Pedestrian detectionFace detectionLocate objects in satellite images (roads, forests, crops, etc.)….
Recognition Tasks Face recognitionFingerprint recognitionIris recognition
…..Video surveillanceCompression (Video, image)Traffic control systems…..
Application Domains
3
1 23 4
2C,s
2s ),(2-(1)2ln(2
)²-(yy)(x,
ttsx
Ss x
x xxTs
s
s
y)(x,minarg Xx
x
Y: Observed Image
X: Hidden Image
HMRF provides an elegant way to model the segmentation problem
This elegant model leads to an optimization problem
Our new proposed approach based on PSO
We looking for The Hidden Image where :
Hidden Markov Random Filed
4
Particles Swarm Optimization Particle related to bird flocking or fish schooling
what's the strategy to find the food?
A group of Particles are randomly searching food in an area
Each particle has a velocity and position
The next position of each particle is influenced by:
The best position, visited by itselfThe best position, visited by all particles
5
Particles Swarm Optimization Advantages
Very efficient global search algorithmSimple implementationEasily parallelized for concurrent processing
Disadvantages
How to choose parameters ?
6
HMRF-PSO methodExperimental Results-(1)
Our tests based on Non Destructive Testing (NDT) DataSet
Examples of NDT application:
Ultrasonic inspection of defective aircraft materials such as
carbon fiber reinforced (CFRP) composites
Thermal inspection of glass-fiber reinforced (GFRP)
Eddy current inspection of aircraft wheel fuselage cracks
Inspection of coating depth of steel plates
Observation of surface roughness of metals and ceramics
Defects in printed circuit board images
In textile image7
(a) (b) (c) (d) (e) (f)
Here some NDT images to segment used in our tests
HMRF-PSO methodExperimental Results-(2)
8
(a) (b) (c) (d) (e) (f)
Ground truth images (original images) of the images listed before
(a) (b) (c) (d) (e) (f)
Here we list the segmented images (tested images) using HMRF-PSO for the parameters:size=80, c1=0.7, c2=0.8, w=0.7, vmax=5, iteration_number=100 and B=1
HMRF-PSO methodExperimental Results-(3)
9
Misclassification Error (ME) ME is used to evaluate the quality of segmentation
ME =0 The best case ME=1 The worst case
F
FF1ME
BO and FO denote the background and foreground of the original (ground-truth) imageBT and FT denote background and foreground of the segmented image
HMRF-PSO methodExperimental Results-(4)
10
HMRF-PSO methodExperimental Results-(5)
Method Image (a) Image (b) Image (c) Image (d) Image (e) Image (f)Abutaleb 0.023 0.310 0.023 0.024 0.250 0.620Kittler-Ill. 0.000 0.003 0.037 0.008 0.025 0.028Kapur et al. 0.003 0.004 0.028 0.036 0.220 0.620Tsai 0.240 0.170 0.350 0.290 0.084 0.280Li & Lee 0.490 0.550 0.450 0.710 0.021 0.020Pham 0.460 0.560 0.021 0.760 0.048 0.250SemiV 0.003 0.004 0.026 0.018 0.062 0.160Otsu 0.462 0.513 0.413 0.021 0.037 0.074Median extension 0.462 0.527 0.474 0.608 0.028 0.039MoG 0.000 0.000 0.032 0.010 0.018 0.012MoGG 0.000 0.000 0.028 0.007 0.012 0.016HMRF-PSO 0.000 0.000 0.018 0.001 0.004 0.005
Misclassification errors in NDT segmented images means the minimal error
11
11
Conclusionwe have presented HMRF-PSO method and compared it to thresholding methods
Performance evaluation was conducted on NDT image dataset
Misclassification Error criterion was used as a performance metric
From the results obtained, the HMRF-PSO combination method outperforms thresholding methods
HMRF-PSO method demonstrates its robustness and resistance to noise
Selecting parameters still not obvious task
12
Thank you for your attention
