Image retrieval using contourlet based interest points

IMAGE RETRIEVAL USING CONTOURLET BASED INTEREST POINTS

Hoang Ng-Duc(1), Thuong Le-Tien(2), Tuan Do-Hong(2), Cao Bui-Thu(3), Ty Ng-Xuan(4)

(1) Broadcast Research and Application Center, Vietnam Television (VTV-BRAC)

(2) Ho Chi Minh City University of Technology

(3) Ho Chi Minh City University of Industry

(4) Navy Institute Nha Trang

Email: [email protected], [email protected], [email protected], [email protected], [email protected].

ABSTRACT

In this work we present the method for image retrieval

based on the Non-Subsampled Contourlet Transform

(NSCT) and the Harris corner detector. The NTSC-based

interest point detector is proposed by combination of

NTSC and Harris corner detector called the Contourlet

Harris detector. We also present the method how to

extract the image features using this Contourlet Harris

detector that is applied for image retrieval. Experiments

are implemented on the WANG database aiming to

compare retrieval effectiveness of proposed method to

some methods have announced. Results demonstrate that

the proposed method shows a quite improvement in the

retrieval effectiveness.

Keywords: content-based image retrieval, CBIR, Non￾Subsampled Contourlet Transform, Contourlet Harris

detector

1. INTRODUCTION

CONTENT-BASED IMAGE RERTIEVAL (CBIR) becomes a

real demand for storage and retrieval of images in digital

image libraries and other multimedia databases.

Basically, CBIR is an automatic process for searching

relevant images to a given query image based on the

primitive low-level image features such as color, texture,

shape, and spatial layout.

References [11], [12] introduced the contourlet

transform with improved characteristics compared with

the wavelet transform. The contourlet transform was

designed by a discrete-domain multiresolution and

multidirection expension using non-separable filter

banks, in much the same way that wavelets were derived

from filter banks. Image decomposition using this

transform has a flexible multiresolution, and local and

directional expansion for images. Many applications have

been developed based on contourlet transform as image

denoising, enhancement, and CBIR. Reference [4]

introduced a CBIR solution based on contourlet transform

with the results is quite good.

Interest points represent local properties of the image

which have local variation in at least two directions. The

main desirable property for an interest point detector is

the repeatability: whether or not the same feature will be

detected in two different images of the same scene.

Several methods have been proposed for interest point

detection, which can be categorized into edge based

methods, intensity based methods, biologically inspired

methods, model based methods, multi-scale methods,

scale invariant methods and affine invariant methods

[13].

The Harris corner detector looks for corners where the

gradient changes in two directions. The direction of

gradient does not affect the detection thus making the

detection rotation invariant. The detector is somewhat

robust to variations in illumination since gradient is

insensitive to changes in brightness. Another well known

detector is Difference of Gaussian (DoG) which is

adopted in Scale Invariant Feature Transform (SIFT) [6]

that also is an algorithm to detect and describe local

features in images.

1.1. Our Approach

In this paper, we propose a new detector to capture

interest points in images called the Contourlet Harris

detector and design a corresponding descriptor for image

features. The highlights of this approach are: (i) it used

Nonsubsampled Contourlet transform with fully shift￾invariant, multiscale, and multidirection expansion that

has improved characteristics compared with contourlet

transform is not shift-invariant due to downsamplers and

upsamplers present in both the Laplacian pyramid and the

DFB (directional flter bank) [2]; (ii) it used the Harris

corner detector [3] extract interest points on subbands of

NTSC; (iii) the image features are computed from

detected interest points and energy of each subband. Our

experiments show that proposed method can outperform

methods based on individual contourlet [4] and NSCT,

coocurrence [14] features for image retrieval.

1.2. Related Works

The use of interest points in content-based image

retrieval allows image index based on local properties of

image. A salient detector that extract points where

10th International Conference on Information Science, Signal Processing and their Applications (ISSPA 2010) 978-1-4244-7166-9/10/$26.00 ©2010 IEEE 93