You are on page 1of 3

Research on Clothing Image Retrieval Technology

and System

Wang Hai-long Du Jun-li


School of Computer Science School of Computer Science
Zhongyuan University of Technology Zhengzhou China
Zhengzhou China e-mail: lududi@126.com
e-mail: whlseadragon@sina.com Zhongyuan University of Technology
AbstractContent Based Image Retrieval is a current research
results, that is, decreasing the scope of search results until the
focus at home and abroad. With the increasing popularity of
required goal of positioning.
clothing e-commerce, the requirements of Image based clothing
retrieval are increasingly urgent. The clothing style is the z The actual image database is not only huge, but the type
essential characteristics. The paper presents the method of and quantity is wide. It requires CBIR system can
contour feature extraction, expression and matching to achieve the retrieval speed as conventional information
implement clothing image retrieval comprehensively. retrieval system.
Experiments show that there methods can achieve the higher
rate of full and accuracy. z CBIR system should be highly interactive, that is, the
user should be able to participate in the process of search
Keywords-Key Words: Clothing image retrieval, Clothing style in order to obtain his desired images.
characteristic, hierarchical search strategy , A typical the basic framework of CBIR system is shown in
Fig.1. As can be seen from the Fig.1, the system can be
I. CONTENT BASED IMAGE RETRIEVAL divided into the image characteristics database establishment
modules and the image database retrieval module [1].
With the development and popularization of digital
technology, computer and network technology More and Image characteristics database establishment modules are
more people have access to a large number of image mainly used to establish and maintain the entire image library,
information. The effective organization, management and its core is the image feature extraction techniques, which
retrieval of large-scale image database have become the takes the image specific characteristics to generate feature
urgent issue need to resolve, its the research content of vector, and store them to form the image characteristics
content based image retrieval. database.

The traditional database retrieval use keywords or Image retrieval module is mainly responsible for the
descriptive text to search, which cant meet the needs of the interaction with the user, evaluating the similarity between
people with the network technology development. sample image and the images in library. When the user
submitted sample image, which is analyzed to extract of
The so-called content based image retrieval finds target feature vector, in accordance with the similarity measure
images from the image library in accordance with the specific guidelines, retrieval module give the search results.
content features of media images, its purpose is not to
understand or identify the target image. Its concern is the
rapid discovery of image content information; the user can
search his desired images from the image library in the
acceptable response time
Content-based image retrieval system has the following
characteristics:
z Breaking the traditional retrieval limitations based on
the expression, CBIR extracts content features from the
image, using the content characteristics to index image.
CBIR is a kind of approximate matching. Because of the
inaccuracy of the image content expression, it must be a Figure 1. Image Retrieval System structure
similar search. Therefore, similarity matching in CBIR
system can use stepwise refinement method to get query
Scientific and technological project of Henan Province
(No.0624220102)

978-1-4244-5273-6/09/$26.00 2009 IEEE


II. USING CBIR TECHNOLOGY TO RETRIEVE CLOTHING
With the rapid development of clothing e-commerce,
online apparel shopping is becoming a way of shopping, and
can not be ignored. With the increase of online apparel goods,
the process of purchasing goods with satisfaction is like a
needle in a haystack. When customers have a clear choice,
they often want to shop around.
Clothing enterprises website is the extern window of
promotion and sale, should also provide users tools to
facilitate efficiently to search clothing information.
In the Internet people need an effective search tool for the
garment to understand the latest clothing fashion, designers
can also use this tool to help their design inspiration. This
demand requires CBIR system to meet.
Therefore, studying clothing image retrieval, to provide
users with information in line with their wishes, is very
meaningful.
At present, these applications in the apparel retrieval are
text based, and the text is difficult to express the rich content
of clothing image, it needs an apparel retrieval tool which is
based on the image content, to help user from a large number
of clothing find clothing images desired.
Now there is a wide range of image retrieval technology, Figure 2. Hierarchical search strategy
however, how to rapid and accurate clothing image
retrieval using the characteristics of the garment is the focus a) Determining the clothing category according to
of the study. keywords index Clothing has often classified information,
such as the Men / Women, up/download, the spring/summer/
Clothing style is an essential feature of clothing image, but winter clothes and so on, simple inspection related classified
also the characteristic is most difficult to describe. In practice,
information determine whether the sample image and the
people will care about the overall style of clothing, may also
retrieve images are similar, give up the different category
be interested in some partial style, such as collar, sleeve.
Therefore, there are overall style image retrieval and local clothing, which is no longer dealt with.
image retrieval. b) Determining the image aspect ratios with the
bounding box. The above similar clothing color images are
Since the system is designed to help users extract converted to grayscale image, determine its bounding box[3],
information related apparel products, using recall factor as the and calculate its aspect ratio. Compared the aspect ratio of
key indicators.
bounding box between sample clothing image and the above
III. CLOTHING STYLES BASED IMAGE RETRIEVAL similar image, if they are approximate, and then continue,
otherwise give up due to different clothing styles.
Sample image should be pre-processed. First image c) Normalization and collimation. The above similar
clothing edge is extracted by using canny algorithm, and then clothing images are dealt with scale transformation to get the
redundant edge is eliminated with the help of same size with sample image, convenient to follow the
human-computer interaction to obtain the styles models of contour extraction.
sample images [2].
d) The outside contour feature extraction and matching:
A. The overall clothing style based image retrieval using canny algorithm to extract the clothing image edge,
determine the perpendicular bisector, from top to bottom
Taking into account the real-time requirements of retrieval
and the time complexity of feature extraction and matching, sub-100 scanning lines to scan ,calculated the distance
hierarchical search strategy is presented, from coarse-to-fine between the left contours, the right contours and perpendicular
step by step, shown in Fig. 2. First simple inspection excludes bisector, store them in array as the outline characteristics.
large deviation quickly, using image profile feature extraction Subtracting two arrays and accumulating its difference. One
and matching to get the retrieval accuracy. array is the feature of sample image, the other is the feature of
retrieval image in library, from small to big difference the
retrieval result is sorted ,an front 30 image is selected as
similar images.
e) Details feature extraction and matching: Scanning
sample images, for each scan line, each pixel in a small
outline, which coordinates is (x, y), find the corresponding
pixel in the retrieval image, cumulate its deviation. After the
scan is complete, the deviation accumulator is the distance
between example image and retrieval images. from small to
large deviation accumulator the retrieval result is sorted.
Fig.3 is given to show the overall clothing style based
image retrieval results.

Figure 4. The partial clothing style based image retrieval results

IV. CONCLUSION
In this paper, the retrieval method applies to the ideal
one-piece garment retrieval. The experimental show that the
retrieval results can satisfy However, in practical applications,
images are often model wearing clothes, clothes hanging in
Figure3. The overall clothing style based image retrieval results the racks or local folding clothes. Therefore, further work is
carried out on these complex images pre-processing research,
B. The partial clothing style based image retrieval in order to adapt to a wider range of applications.
z First of all, the user determines the interested area in the REFERENCES
sample images, to have an interested image block [4, 5].
[1] Q. Xu, W.W. Yang , S.T. Chen Content - Based Image Retrieval
z Using the global type feature extraction methods to Technology(in Chinese) Computer Technology and Development
obtain the image block characteristics named K1. 2008 18(1):126-128
[2] Y.Q. Sun, S.J. Ozawa. A Hierarchical Approach for Region-Based
z using Canny algorithm to obtain target image region of Image Retrieval. IEEE. 2004: 11171124
retrieve images ,which is named K2. [3] Y. Zhang, S.L. Ma, X.Han, Z.B. Zhang Image Extraction and
Segment Arithmetic of License Plates Recognition(in Chinese).
z In region K2 search the region which is similar to K1 in Journal of JiLin University(Science Edition)2006 44(3):407-408
style feature. [4] G.Zhai, Y.Q. Song, J.N.Tan, Q.Xu Image retrieval based on
combining interest points and edges for content(in Chinese).
z To speed up the search speed, according the position of Computer Engineering and Design 2009,30 (5):1148-1150
interested block in the sample image, in the retrieval
[5] FJMeng,B.L.Guo,A novel image retrieval algorithm based on the
image we use the corresponding region, which is color and distribution of interest points(in Chinese).  Journal of
expanded about 20% outward as search scope. XiDian University, 2005,32(2):256-259
Fig.4 is given to show the partial clothing style based
image retrieval results.

You might also like