An Approach to Cbir using K - means Clustering and Polygon based shape features

Dominic Mai Tạp chí KHOA HỌC & CÔNG NGHỆ 64(02): 45 - 52

45

Số hóa bởi Trung tâm Học liệu – Đại học Thái Nguyên http://www.Lrc-tnu.edu.vn

AN APPROACH TO CBIR USING K-MEANS CLUSTERING

AND POLYGON BASED SHAPE FEATURES

Dominic Mai, Toi Nguyen Van

Faculty of information Technology Thai Nguyen University

ABSTRACT

In this paper an approach to Content Based Image Retrieval (CBIR) is examined that uses K￾means clustering for segmenting an image and then extracts global and local features using

color and shape information over the extracted regions to compute a similarity measure between

images. Although color is used as main information source for creating the clusters that an image

is formed of, shape factors of the separated regions will be taken into account for the retrieval

process as color is heavily dependent on the lighting of a scene. A fuzzy representation of

features is chosen that suits the blurry nature of image segmentation.

Từ khóa: Shape information, K-means, global feature, local feature, segment an image

*

INTRODUCTION

In 2006 over 300 million photos were

uploaded to flickr, one of the biggest photo

sharing communities on the internet. This

number is just for illustrating the fact that the

number of pictures stored in electronic

databases is increasing rapidly and the task of

efficiently retrieving pictures stored in such a

database is becoming more and more

important. Text based search techniques can

only be applied if pictures have been assigned

meaningful labels describing semantic entities

like people, outdoor scene, etc.

Unfortunately, understanding a picture in the

way humans do is a very hard task that is not

yet solved by automated algorithms – this is

also known as the semantic gap.

As most of pictures do not come labeled or

inside a labeled context (e.g. a website)

automatic retrieval of images cannot be done

using text retrieval techniques.

Luckily, it is not necessary to semantically

understand a picture for performing a

satisfying retrieval on a database. In this

paper an approach to Content Based Image

Retrieval (CBIR) is examined that uses K￾means clustering for segmenting an image

and then extracts global and local features

using color and shape information over the

extracted regions to compute a similarity

measure between images.

Although color is used as main information

source for creating the clusters that an image

is formed of, it will not play a major role

*

Tel: 0912847077

for the retrieval process as the concept of

color is quite difficult [1] – just think of the

different white balance settings of your

camera you have to make when shooting

under different lighting conditions to make

white appear as white. Therefore the shape

of the found regions is analyzed using a

polygonal representation of the shape’s

boundary which is not dependent on color

information.

To account for the inherent uncertainty

coming with the applied segmentation

process – a good segmentation cannot be

achieved without semantically understanding

the image – a fuzzy representation of the

regions is applied.

A reference System has been implemented

using JAVA.

RELATED WORK

This paper is mainly based on the ideas for

segmentation and similarity computation

presented in [3]. In the process of analyzing

the image, it is partitioned into blocks of

4x4 which contain average color values

over the three bands in the RGB color

model. These blocks (Blockfeatures) are then

clustered using an iterative K-means [2] with

a small refinement that leads to superior

performance over randomly choosing the

initial point set.

A Cauchy membership function is used to

account for fuzziness in assigning

Blockfeatures to their respective clusters. It

turns out that only the center points of a

cluster are needed for the comparison

measure applied which leads to very little