Color Clustering and Learning for Image Segmentation

Color Clustering and
Learning for Image
Segmentation Based on
Neural Networks
2009.02.23
최지혜
Guo Dong, Member, IEEE, and Ming Xie, Member, IEEE
Topic
신경망을 기초로한 영상분할을 위한 컬
러 클러스터링과 학습
주제어 : SOM (능동적으로 맵을 구성)
SA (전역 최적 클러스터링 위해)
HPL(계층적 초기학습)
요약
• 정확한 색상차이를 측정하기 위하여 수정
된 컬러공간 L*u*v*를 사용하였다. 분할
시스템은 supervised/unsupervised
segmentation으로 이루어 졌는데, un의 경우 컬러감소와 컬러 클러스터링을 목
표로 할 수 있다. 컬러 클러스터링은
SOM-SA의 장점들로 취해 저 연산비용으
로 최적의 분할을 이룰 수 있다. HPL은 색
상 프로토타입들을 목적색상을 위해 좋은
근사처리로 공급된다.
Image segmentation system
based on neural networks
미리 알기 어려운
이미지
오브젝트의 컬러를
알고 있는 경우
Unsupervised segmentation
미리 알기 어려운
이미지
Spatial Compactness
Color homogeneity
Desirable properties
Image-domain
Feature-space
Segmentation techniques
Unsupervised segmentation
Splitting and merging phases
Watershed transform
Self-organizing map
SOM
SOM
• SOM is trained to generate the
primitive clustering
dominant colors of the image
Description of Problem
• To ensure a proper measure of color
differences, image colors must be
represented in a uniform color space.
• In unsupervised segmentation, color
reduction is indispensable to the
segmentation of a large color image.
• In supervised segmentation, color
learning is crucial to build up an
accurate classifier for the segmentation
of the object of interest.
Flow of This paper
Appropriate color space
color reduction
is performed by SOM learning
SA seeks the optimal clusters from
SOM prototypes
New procedure of supervised learning
Appropriate color space
L* is luminant component
u* and v* are color components : u* axis varies from green to red
v* axis changes from blue to yellow
RGB to L*u*v*
Color reduction : SOM learning
a two-layer neural network with a
rectangular topology
Three inputs are fully connected to the neurons on a 2-D plane.
Each neuron is a cell containing a weight values.
SOM Training
• Initialization – 16x16 rectangular
neighborhood type is Gaussian
weight vector –randomly initialize
radius r = 16, 5 learning rate = 0.05,0.02
• Input - each color point
• Competitive Process – ‘wining neuron’
SOM Training
• Cooperative Process- The topological
neighbors are determined by
Gaussian function centered at
• Adaptive Process - The weights of
“winning neuron”
and its neighbor
neurons are updated within the
neighborhood
: Effective scope
: neighborhood
function
Sammon mapping of 16x16 weight
vectors after SOM training
SA seeks the optimal clusters from
SOM prototypes
Simulated annealing은 커다란 탐색공간에서 주어진 함수의 전역
최적점에 대한 훌륭한 근사치를 찾으려고 하는 전역 최적화 문제에
대한 일반적인 확률적 휴리스틱 접근방법
고체의 물리적인 담금질과 아주 많은 경우의 수를 가진 조합최적화문
제사이의 밀접한 관계-> 여러 다른 신경망의 학습과정을 변화시켜줄
수 있다.
학습한다 :minimization 과정으로 볼 수
있으며,
이것은 energy function이나 error
function에서 downward 방향으로 간다.
Initial weight 잘못 선택 시
Local minimum
SA의 개념의 도입
SA
• The optimal solution is obtained by
consisting in randomly perturbing the
system, and gradually decreasing the
randomness to a low final level.
cluster centers be
The criterion of sum-of-squared-error
The procedure of SA clustering is to search the appropriate
cluster centers = minimize the energy function
SA Clustering
Clustering
Segmentation result by
SOM-SA color clustering
New procedure of supervised learning
• RCE neural network is a supervised
pattern classifier used for the
estimation of feature region
• hyperspherical window
Drawback of RCE learning
• requirement of a complete sample
set for all classes
• it requires the samples of both the
object and the image background.
– to segment the object of interest from
the image background
Hierarchical Prototype Learning
• In some regions, a small size of
prototype is appropriate,
• Other regions, a large size of
prototype is more suitable.
>> The proper way of region
estimation is to estimate the region
by the different sizes of color
prototypes.
Hierarchical Prototype Learning
EXPERIMENTAL EVALUATIONS
(a)Original color image. (b) SOM color clustering. Q =5580.824.
(c) SA clustering. Q = 182.526.
(d) SOM-SA color clustering. Q =244.826.
(e) CL-SA color clustering. Q = 376.845.
Supervised segmentation
(a) Original gesture image. (b) HPL learning. (c) Color threshold.
(d) Color histogram.
(a) Original hand gesture images. (b) Segmentation of hand gestures.