Multi-feature Content-based Product Image Retrieval
Based on Region of Main Object
Lunshao Chai
1
Honggang Zhang
1
Zhen Qin
2
Jie Yu
1
Yonggang Qi
1
1
Pattern Recognition and Intelligent System Laboratory
Beijing University of Posts and Telecommunications
Beijing, China
2
Riverside Lab for Artificial Intelligence Research
University of California, Riverside
Riverside, CA, USA.
AbstractContent-based image retrieval (CBIR) has got an
intense interest and seen considerable progress over the last
decade. But most of the time it is only applied in laboratory. One
important reason for this is the diversity of images. Different
practical situations call for different taxonomy definitions of
images, and lead to very different solutions. At present, and even
in the foreseeable future, a general purpose CBIR system is not
really possible. However, image search engines oriented at
specific domains are feasible in technology and also have the
actual demand. With the rapid development of electronic
commerce, searching specific product by image has become one
of the most attractive related research topics. In this paper, we
propose a region-based method fit for the content-based retrieval
of product images. The method focuses on two key issues: fast
extraction of the main region, in which the product locates, as
well as efficient shape and color features extraction. To show the
validity of the proposed region-based method, compared
experiments are carried out and illustrated on the PI 100 dataset.
Keywords-content-based product image retrieval; fuzzy color
histogram; radial-harmonic-fourier moments (RHFMs); region of
interest (ROI)
I. INTRODUCTION
Ever since the day CBIR emerged, it has been considered as
a gold mine with the potential to change the way people search
information. Using text for image retrieval is a mature
technology, but in many cases, it is hard to get precise semantic
description of what people search for. However, an image can
depict an object or a scene clearly and roundly, and with CBIR,
we can avoid information loss and distortion occurring during
describing images. Although with a bright future, general
purpose CBIR systems are still far from being utilized in real
world applications. Nevertheless, we can expect CBIR to
become practical in certain contexts of use in the near future.
Content-based product image retrieval refers to image-to-image
retrieval of product pictures, especially in the case of online
shopping. It is a lively field of research driven by huge demand
and most likely to make a breakthrough.
CBIR of product images is technically possible largely due
to the simplification of segmentation. Even though
segmentation has always been a hotspot of study and new
achievements deliver every year as in [1][2][3], extracting
semantic ROI automatically and precisely is still beyond the
reach of current computer vision techniques. However, there
are some favorable restrictions in product images: background
are rarely complex, and the main object - the product, will be
distinct and composes the center of the image. These
restrictions greatly simplify the segmentation and therefore
image analysis is possible.
In our work, Radial-Harmonic-Fourier Moments (RHFMs)
is adopted to describe the shape of the product. Shape is one of
the most important visual features, in that perceiving and
distinguishing the shape of objects in an image are crucial for
people to understand it. Comparing with lower level features -
color and texture, shape belongs to middle level features, which
are important components to describe high level visual features
like target or object. Moment invariants are one kind of image
features that can satisfy shifting, scaling, rotation, and intensity
invariance. From the first time as moment invariants was
introduced in [4], a number of variants for moments invariants
have been proposed as in [5][6][7][8] and used to describe
images [9][10][11]. RHFMs was first proposed in [12]. The
analysis performed showed that RHFMs outperformed other
known moments in image description, image reconstruction,
and noise-resistance power, especially for small-size images.
Fuzzy Color Histogram is used to describe the color feature
of the image in our method. The technique of color histogram
has been widely used because of its simplification and
resistance to distortions such as rotation and scaling [13][14].
The fuzzy linking technique we choose is proposed in [15] as a
component of Color and Edge Directivity Descriptor (CEDD).
With low computational power needed, the system can
generate an accurate and robust 24-bin color histogram
representing 24 colors in an image.
The rest of the paper is organized as follows. In Section II,
we propose a fast extraction algorithm fit to product images.
Using prior information, this algorithm can extract the main
region, which contains the product, with much less time
expense than traditional segmentation algorithms. In section III,
the main region we extracted is used as region of interest (ROI).
Based on the ROI, shape feature (Radial Harmonic Fourier
Moments, RHFMs) and color feature (fuzzy color histogram)
are then extracted. In section IV, we compare the performance
of features extracted with our region-based method with four
other commonly employed non-region-based features. Section
V concludes and future work is discussed.
This work was partially supported by National Natural Science Foundation of
China under Grant No.61005004, the Fundamental Research Funds for the
Central Universities and Scientific Research Foundation for the Returned
Overseas Chinese Scholars, State Education Ministry. This work is also
funded by Qualcomm, Inc.
978-1-4577-0031-6/11/$26.00 ©2011 TEEE TCTCS 2011