1 Multiple Classifier Based on Fuzzy C-Means for a Flower Image Retrieval Keita Fukuda, Tetsuya Takiguchi, Yasuo Ariki Graduate School of Engineering,

1

Multiple Classifier Based on Fuzzy C-Means for a Flower Image Retrieval

Keita Fukuda, Tetsuya Takiguchi, Yasuo ArikiGraduate School of Engineering, Kobe University, Japan

2

• Introduction– Purpose of Multiple Classifier based on Fuzzy C-means

– Overview of our flower recognition system

• Proposed Method– Each Classifier

– Fuzzy C-means

• Experiments

• Summary and Future Work

Table of contentsTable of contents

3

IntroductionIntroduction

What is this flower ?Retrieval system requires “keywords.” But it is difficult to get “keywords” from “images.”

We take a flower picture and send it to a system. We receive flower image information there and then immediately.

We are focusing on flower image retrieval system

In our proposed method

4

Conventional techniquesConventional techniques

Conventional method• Using the same features for classification.⇒But flowers have various shape.

• We propose multiple classifier which selects important features for each flower type and weights the importance on each classifier using Fuzzy c-means.

It is required to select important features according to flower type.

5

Overview of our systemOverview of our system

Send image

Receive information

Database

Flower region extraction

Color and shape features extraction

Similarity by multiple classifier

contents based flower image retrieval

6

Flower region extractionFlower region extraction

A large color regions locating at near center are extracted as flower region

Color and Shape features are computed on them.

7

Feature extractionFeature extraction

Color feature

0.42

0.03

0.00

10

10

Distribution histogram

Shape feature

d

ld

Power

Freq

Power spectrum

Fourier transform

l: contour pixel

d: distance from G to contour

Gravity to contour

100 segments

G

8

Similarity for each classifier is calculated

Recognition with multiple classifier (1)Recognition with multiple classifier (1)

Query image

Image

Multiple classifier

BF

CF

AF 0.03

0.93

0.04

iAV

iBV

iCV

Similarity for classifier

Membership of query image

(Weight)

)(iM

i

jF

Similarity

Features, Information, Similarity

…Database

We define 3 classifiers for 3 flower types

Membership of a query image in each type is obtained as weight for each similarity

Linearly coupled similarity matching of 3 classifiers

Which types is a query image

associated with ?

9

Each flower typeEach flower type

Type Classifier For Similarity

A FA “Near circle” flowers ViA

B FB “Clear one petal” flowers ViB

C FC “many petals” flowers ViC

The similarity in each classifier is computed using Weighted Histogram Intersection. The value of weight represents the

difference of each classifier

We define 3 classifiers for 3 flower types:

A: “Near circle” B: “Clear one petal” C: “Many petals”

10

Each ClassifierEach Classifier

Gaussian Weight

Histogram Intersection

Query image

image i

In type AIn type BCharacteristics

Peak (5) = the number of petal

Weight of low frequency rangeWeight of band frequency range based on peak

Important similarity

In type C

Weight of high frequency range

11

Query image

Image

Multiple classifier

BF

CF

AF 0.03

0.93

0.04

iAV

iBV

iCV

Similarity for classifier

)(iM

i

jF

Similarity

Features, Information, Similarity

…Database

Membership of query image

(Weight)

Weight for each similarity is membership of a query image in type A, B and C

⇒ It is difficult that all flowers are classified into one of 3 types clearly.

Recognition with multiple classifier (2)Recognition with multiple classifier (2)

12

Fuzzy C-meansFuzzy C-means

N

i

C

jji

mijm cxuJ

1 1

2

It is based on minimization of the following objective function:

Fuzzy partitioning is carried out through an iterative optimization of the objective function, with the update of membership uij and the cluster centers cj

N

iijij uiallforu

1

1,0,1

Membership property is

Data elements can belong to more than one cluster. associated with each element is a set of membership.

13

Fuzzy C-means For flower retrieval systemFuzzy C-means For flower retrieval system

1. Database images are clustered using fuzzy c-means.

2. Membership of a query image is computed.

membership

{0.88, 0.05, 0.08}

Data elements

{0.43, 0.41, 0.12}

{0.08, 0.12, 0.80}

C: ”Many petals”

B: ”Clear one petal”

A: ”Near Circle”

Membership of a query image is obtained as weight for each similarity {0.03, 0.93, 0.04}

Input data: shape features {compactness, entropy, average} Output data: membership of the image in each type.

14

Query image

Image

Multiple classifier

BF

CF

AF 0.03

0.93

0.04

iAV

iBV

iCV

Similarity for classifier Membership

(Weight)

)(iM

i

jF

Similarity

Features, Information, Similarity

…Database

iCiBiA VVViM 04.093.003.0)(

Linearly coupled similarity matching of 3 classifiers is calculated. This example, the similarity between image “i” and a query image:

Recognition with multiple classifier (3)Recognition with multiple classifier (3)

15

Result informationResult information

Result information are shown to users up to fifth rank based on the similarity M(i)

Input image Result information

16

Experimental conditionExperimental condition

• Flower images of 120 species with each 4 samples.

(i.e. 480 images in total).

Four Cross validation (evaluate : cumulative recognition)

• One sample is used as a query image (120).

• The others are used as the database images (120×3).

17

Conventional methodConventional method

• Compactness

• The number of petal (peak)

• Moment

• The ratio of the shortest width over the longest

Largest segment

• X coordinate

• Y coordinate

• Its distributed value

2nd Largest segment

• X coordinate

• Y coordinate

• Its distributed value

Shape features Color featuresy

x

peak

18

Experimental resultExperimental result

1st 3rd 5th 10thConventional method 33.8 59.6 69.4 80.8

Multiple classifier

No fuzzy 39.8 67.1 78.1 89.4fuzzy 42.7 69.6 81.3 92.5

Proposed method Conventional methodquery

19

New concept:multiple classifier which select important features for each flower type

SummarySummary

Multiple Classifier Based on Fuzzy C-Means for a Flower Image Retrieval

In future work: research for more than three classifiers

20

Thank youThank you