My2421322135

A. Jwalitha, M.V.Srikanth / International Journal of Engineering Research and Applications
(IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 2, Issue4, July-August 2012, pp.2132-2135
Blurred And Compressed Trademark Image Retreival Under
Noise And Orientation Based On Curvature Shape Descriptor
(Csd)
A. Jwalitha1 , M.V.Srikanth2
1
Student II M.TECH
2
Assistant Professor, Department of ECE
Gudlavalleru Engineering College, Krishna (Dt), India, A.P.

Abstract
Digital images are a convenient media for distance is shown to handle both complete and partial
describing and storing spatial, temporal, spectral, multi-textured queries. But, there are several
and physical components of information contained problems with these methods. First of all, human
in a variety of domains (e.g.. aerial/satellite images intervention is required to describe and tag the
in remote sensing, medical images in telemedicine, contents of the images in terms of a selected set of
fingerprints in forensics, museum collections in captions and keywords. In most of the images there
art history, and registration of trademarks and are several objects that could be referenced, each
logos). Retrieval of digital images is one of the having its own set of attributes. Further, we need to
challenging issue in any Digital Image Processing express the spatial relationships among the various
system. Although advances in image compression objects in an image to understand its content. As the
algorithms have alleviated the storage size of the image databases grow, the use of
requirement to some extent, the large volume of keywords becomes not only complex but also
these images makes it difficult for a user to browse inadequate to represent the image content. The
through the entire database. Therefore, an keywords are inherently subjective and not unique.
efficient and automatic procedure based on Often, the preselected keywords in a given
curvature shape descriptors is proposed, which application are context dependent and do not allow
make use shape descriptor along with for any unanticipated search. If the image database is
compression techniques to make database feasible to be shared globally then the linguistic barriers will
to store large number of images and retrieve make the use of keywords ineffective.
image under noise, blurrring and orientation Color is one of the most widely used low-
changes. In CSD approach, the image is subjected level features in the context of indexing and retrieval
to compression and then later it is represented in based on image content [7]. It is relatively robust to
its contour format by its coordinates and are background complication and independent of image
mathematically processed for curvature evolution size and orientation. Typically, the color of an image
over various sigma levels so as to remove the is represented through some color model. One
unevenness caused by some external disturbances. representation of color content of the image is by
For each sigma level, zero crossing points are using color histogram. For image retrieval, histogram
evaluated which are used as the features for image of query image is then matched against histogram of
retrieval along with arc length. all images in the database using some similarity
metric. However, color histograms do not incorporate
Keywords: Arc length, Blurring, Compression spatial adjacency of pixels in the image and may lead
technique, Curvature Shape Descriptor (CSD), to inaccuracies in the retrieval.
Noise, Orientation changes, Zero crossing points Although color can be an effective means of
querying, color alone as a retrieval cue cannot be
1. Introduction effective for querying large image databases.
Colour, texture and shape information have Applications with grayscale or binary images have to
been the primitive image descriptors in content based use other cues such as shape and texture for retrieval.
image retrieval systems[3]. Traditionally, textual Although, humans can effectively use color to
features, such as filenames, caption and keywords differentiate among natural objects, many artificial
have been used to annotate and retrieve images. (manmade) objects cannot be distinguished on the
Nonlinear modified discrete Radon transform [5] has basis of color alone. Moreover, humans when
been applied to estimate visual contents of textural presented with binary or grayscale images can easily
information of an image, such as orientation, distinguish among these.
directionality, and regularity. Images of either Image description consists in one of the key
individual or multiple textures are best described by elements of multimedia information description. In
distributions of spatial frequency descriptors, rather the Multimedia Content Description Interface
than single descriptor vectors over presegmented (MPEG-7) images are described by their contents
regions. A retrieval method based on the Earth featured by color, texture and shape. The shape
Movers Distance [6] with an appropriate ground descriptor aims to measure geometric attributes of an

2132 | P a g e

object to be used for classifying, matching, and important frequencies are discarded through
recognizing objects.There are several techniques quantization and important frequencies are used to
available for shape representation as Fourier retrieve the image during decompression.
descriptors, Wavelet descriptors, grid-based, The rest of paper is organized as follows. In
Delaunay triangulation, among others. Shape section 2, we explain the algorithm for proposed
description techniques are classified into boundary method. In section 3, the results of the proposed
based and region based methods. Boundary based system are shown and its performance is analysed.
methods use only the contour of the objects’ shape, Section 4 gives concluding remarks
while the region based methods use the internal
details in addition to the contour. The specified Shape 2.NOVEL APPROACH FOR IMAGE
representation methods proposed fail to satisfy one or RETRIEVING
more of such as criteria Invariance, Uniqueness, This section describes novel approach for
Stability, Efficiency, Ease of implementation, and image retrieving using Curvature Shape Descriptor.
Computation of shape properties. In this approach the trained images are subjected to
compression and then later it is represented in its
The classical Curvature Scale Space method contour format by its coordinates and is
for contour representation captures describes and mathematically processed for curvature evolution
compares characteristic shape features of over various sigma levels so as to remove the
objects[1][2][4]. It represents two dimensional shapes unevenness caused by some external disturbances.
at different resolutions. Maxima (peaks) of CSS For each sigma level, zero crossing points are
images are used to describe the shape and to perform evaluated which are used as the features for image
matching between two curves under analysis. The retrieval along with arc length. Later the retrieval
matching scheme is based on the Euclidean distance module receives user query, applies CSD approach to
between the peaks of the CSS images. This scheme is obtain image features arc length and zero crossing
very complex and expensive points .The retrieval module is based on Euclidean
distance that compares query image features with
High-resolution images can occupy large trained images in the database. The design
amounts of storage (around 17.5 Mb for one A5 color architecture for proposed approach is specified below
image scanned at 600 dpi). The need to compress
image data for machine processing, storage and
transmission was therefore recognized early on. To
overcome these effects specified, CSD approach has
been proposed, which is an efficient and automatic
procedure is required for indexing and retrieving
images from databases. In this approach the image is
subjected to compression and then later it is
represented in its contour format by its coordinates
and is mathematically processed for curvature
evolution over various sigma levels so as to remove
the unevenness caused by some external
disturbances. For each sigma level, zero crossing
points are evaluated which are used as the features
for image retrieval along with arc length.

Data compression is the technique to reduce
the redundancies in data representation in order to
decrease data storage requirements and hence
communication costs [8][9][10]. Reducing the
Fig 2.1 system architecture
storage requirement is equivalent to increasing the
The designed system architecture is as
capacity of the storage medium and hence
presented in figure 2.1 The algorithm for proposed
communication bandwidth. Thus the development of
approach is specified below.
efficient compression techniques will continue to be a
2.1 DCT BASED COMPRESSION High-resolution
design challenge for future communication systems
images can occupy large amounts of storage (around
and advanced multimedia applications. A technique
17.5 Mb for one A5 color image scanned at 600 dpi).
to Image compression is the application of Data
The need to compress image data for machine
compression on digital images. The discrete cosine
processing, storage and transmission was therefore
transform (DCT) is a technique for converting a
will continue to be a design challenge for future
signal into elementary frequency components. It is
communication systems and advanced multimedia
widely used in image compression. DCT separates
applications. A technique to Image compression is
images into parts of different frequencies where less

2133 | P a g e

the application of Data compression on digital P' and Q' are the first order derivative of p, q and P'',
images. The discrete cosine transform (DCT) is a Q’’ are the second order derivatives.
technique widely used in image compression. C (u, σ) is smoothened curvature.
[10]DCT separates images into parts of different For the obtained smoothened curvature at each
frequencies where less important frequencies are gaussian level, zero crossings are computed.
discarded through quantization and important
frequencies are used to retrieve the image during 2.6 ZERO CROSSING COMPUTATION
decompression. Typically, input taken images are After smoothening the given curvature a
compressed in macroblocks of 8 rows by 8 columns, zero cross is evaluated, where the zero cross is found
where each block is linearized into a one-dimensional when the tracing come across a pixel variation from 0
vector. Various level of compression can be achieved to 1 or 1 to 0 level.
using this algorithm. The compressed image is passed
to preprocessing unit (canny edge detection) for 2.7 SHAPE DESCRIPTOR (CSD) EVALUATION
minimising the surrounding effects. Once the zero cross were obtained they are
buffered for a corresponding arc length (u) and given
2.2 CANNY EDGE DETECTION Gaussian value (σ), once all the zero cross were
The compressed and smoothened image is found they are been plotted for arc length v/s sigma.
passed as input for canny edge detection unit. This From the obtained CSD plot important curvature
edge detection is performed using defined canny edge features are extracted. To obtain the important
operators. Once the preprocessing operation is done curvature feature a threshold is set given by
the obtained edge information is passed for CSD Threshold (T) = 0.6 * max. Peak value. This indicates
evaluation, where the initial operation is to evaluate that a CSD peak of more than 60% of the obtained
the contour of the given edge information. curvature information is used for image feature
representation. This approach eliminates the
2.3 BOUNDARY EXTRACTION consideration of lower peaks resulting in elimination
Contour is defined as outermost continuous of shape information generated due to external
bounding region of a given image. For the detection noises. This noise used to be reduced by filtration
of contour evaluation all the true corners should be approach in conventional methods.
detected and no false corners should be detected. All
the corner points should be located for proper 2.8 FEATURE EXTRACTION AND
continuity. The contour evaluator must be effective in RECOGNITION:
estimating the true edges under different noise levels Once the CSD plot is obtained, features are
for robust contour estimation. For the estimation of extracted from the selected CSD peaks, then grouped
the contour region an 8-region neighbourhood- together for matching and classification, given as
growing algorithm, once the contour is detected the Feature={(σ1,arclength1), (σ2, arclength2)…(σn,
curvature for the obtained contour is calculated. arclengthn)},Where n= no. of peaks crossing the
2.4 CURVATURE EVALUATION threshold limit. This feature extraction is done for the
To evaluate the curvature for the obtained given database information and query image which
contour of given image following approach is made. are used for recognition. Once the knowledge is
For a given a contour co-ordinates (p (u), q(u)) the created the recognition operation is performed.
curvature of the given contour is given by
𝑝′ (𝑢 )∗ 𝑞′′ (𝑢 ) – 𝑞′ (𝑢) ∗ 𝑝′′ (𝑢) 3. EXPERMENTAL RESULTS AND
C (u) =
[( p′ (u ))^2 + (q′ (u))^2]^(3/2)
PERFORMANCE ANALYSIS
Where (p’, q') are first derivative of given contour
highboost image2
3.1.RESULTS
co-ordinates and (p’’, q'') are the double derivative of Original Image

p and q.
C (u) is the curvature of the given image
.For the obtained curvature, CSD is obtained by
applying smoothening operation to reduce the zero
crossing co-ordinates in bounding contours. The
smoothening is continued by incrementing the
Gaussian value (σ) on the obtained contour until no
zero crossing exists.
(a) (b)
2.5 CURVATURE SMOOTHENING 20

18
CSS-final Recognized Image
The curvature smoothening is done using equation 16

14

𝑃’ (𝑢 ,𝜎 )∗𝑄’’ (𝑢,𝜎 )−𝑄′ (𝑢,𝜎 )∗𝑃′′ (𝑢 ,𝜎 )
C (u, σ) = [(P′ (u,σ))^2 + (Q′ (u,σ))^2]^(3/2)
12

10

8

Where e, P = conv (p, g) and Q = conv (q, g), 6

4

g is Gaussian distribution function and u is the arc 2

0 5 10 15 20 25 30 35 40 45 50

length parameter (c) (d) (e)

2134 | P a g e

accuracy of CSD high, when compared to Invariant-
Moment based method and also CSD method
requires less retrieval time which is compared to
Invariant-Moment based method for retrieving the
same image shown in figure 3.2.
(f) 4
Comparision of retrieval time factor between CSD and Invarient method

Invr-Method
Invarient based Recognition 3.5 CSD-Method

3

Computation time (sec)
2.5

2

1.5

1

(g) 0.5

Fig 3.1 (a) input image (b) test image subjected to 0
1
Methods
2

compression,blurring and rotation,(c)CSD plot Fig3.2
Comparision of Recogntion Accuracy factor between CSD and Invaient method

(d)classified images in CSD method (e) CSD based 300
Invr-Method
CSD-Method

recognized image (f)classified images in Invarient 250

moment method (g) Invarient method based 200

recognition level
recognition 150

A trademark query image shown in Fig 100

3.1(a) is subjected to blurring and is 25% compressed 50

and rotated at 90 degrees as shown in Fig 3.1(b) 0
1
method
2

which is passed to the algorithm for retrieval purpose. Fig3.3
Based on the features of CSD shown if Fig 3.1(c) and 4. CONCLUSION
the moments (M1 to M7) of Invariant moment An efficient shape-based retrieval algorithm
method, Classification has done and shown in Fig has been developed to retrieve image which have
3.1(d) and Fig 3.1(f) respectively. From Fig 3.1(e) shown to be a promising technique for shape-based
and Fig (g), it can be observed that the CSD method
image database retrieval. The CSD based method is
is again more efficient in retrieving images when it is
compared with the invariant based estimation
blurred and compressed and when subjected to noise
algorithm and observed to be robust under
and orientation changes.
compression, blurring, rotation, and noisy versions of
the database images. Further it is observed that the
3.2 PERFORMANCE RESULTS CSD based estimation outperforms the existing
The Figures given below specifies
techniques with higher accuracy and less
performance analysis between CSD and Invariant classifications for retrieval of an image.
Moment based methods under compression, blurring
and compressed with noise corrupted and subjected
to orientation cases. figure 3.3 shows that retrieving
[5] Mahmoud R. Hejazi, Yo-Sung Ho, An
REFERENCES Efficient Approach to Texture-Based
[1] Sadegh Abbasi, Farzin Mokhtarian, Josef Image Retrieval ,Vol. 17, 295–302
Kittler , Curvature scale space image in (2007), 2008 Wiley Periodicals, Inc.
shape similarity retrieval, Centre for [6] Rishav Chakravarti, Xiannong Meng, A Study
Vision Speech and Signal Processing, of Color Histogram Based Image
Department of Electronic & Electrical Retrieval, Sixth International Conference
Engineering, University of on InformationTechnology2009: New
SurreyGuildford GU2 5XH, UK; Generations, IEEE
[2] Anil K. Jain and Aditya Vailaya, shape- [7]Faisal Bashi, Shashank Khanvilkar, Ashfaq
based retrieval – A case study with Khokhar, and Dan Schonfeld, Multimedia
trademark image databases, Dept of Systems: Content Based Indexing and
computer science, Mitchigan University. Retrieval.
[3] W. Niblack et al, The QBIC project; [8] R.C. Gonzalez and R.E. Woods, Digital
querying images by content using colour, Image Processing (Addison Wesley
texture and shape, in SPIE, VOL. 1908, Publications).
1993. [9] William K. Pratt. Digital Image processing
[4] Yossi Rubner and Carlo Tomasi, Texture- (John Wiley and Sons, Inc, New York,
Based Image Retrieval without NY, second edition, 1991).
Segmentation, Computer Science [10]Anil K. Jain, Fundamentals of Digital Image
Department ,Stanford University, Processing (Prentice-Hall, Englewood
Stanford, CA 94305. Cliffs, NJ, 1989).

2135 | P a g e

My2421322135

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to My2421322135 (20)

My2421322135