International Journal of Emerging
Technology & Research
Volume 1, Issue 4, May-June, 2014
(www.ijetr.org)
ISSN (E): 2347-5900 ISSN (P): 2347-6079
Improving Performance in Retrieval of Images from Multimedia
Database by an Automated Process
Snehaa.R1, Preethi.S2
1, 2
Computer Science and Engineering, Agni College of Engineering, Chennai, TamilNadu, India
Abstract— Nowadays, the Multimedia Databases plays a
vital role in many applications . In medical arena, the
number of new technology is increasing rapidly. A medical
image storage and retrieval system includes database. It is
necessary to store huge amount of data in it. Hence it
makes the retrieval process very slow which leads to
inflexibility and unreliability. It is an automated process
deals with the Filtering, Edge Detection, Sketching, Scale
map Formation, Sub-Segment formation, Corners
Detection. The resultant will be fetched to the Database by
using an index. The Retrieval process has been carried out
in Feature Extraction. The process has been done by using
three ways CBIR, Query by Example (QBE) and Query by
Sketch (QBS). CBIR which is the Content Based Image
Retrieval used to match the Images in the Database. Query
by Example ,which is used to retrieve both shape and
texture descriptors of the images based on edges and
corners. Query by Sketch ,which is used to retrieve the
Images based on Sketched Image.
In order to extract the relevant sub segments, which are
characterized by long, connected series of relatively strong
edge-pixels, from the scale-map as the first step and then a
novel shape description, as referred to 2-D walking ant
histogram (WAH), is applied over them. It is basically
motivated from the following imaginary scenario. A sample
illustration of such a scenario is shown in Figure. 1.
Keywords-Corners and Edge Detection, Query by Example, Query
by Sketch, Sketching.
Figure 1. Walking ant descriptor on a Sketched Image.
I. Introduction
Content-based retrieval uses the contents of multimedia to
represent and index the data. In typical content-based retrieval
systems, the contents of the media in the database are
extracted and described by multi-dimensional feature vectors,
also called descriptors. The feature vectors of the media
constitute a feature dataset. To retrieve desired data, users
submit query examples to the retrieval system. The system
then represents these examples with feature vectors. To
improve the efficiency of the content-based retrieval in
multimedia databases, the relevant sub segment of the shape
descriptor is used for the indexing. In order to find the relevant
sub segment of the image, the exact edge of the image should
be found by using filters and edge detectors along with the
scale map.
© Copyright reserved by IJETR
Suppose an ant is walking over a solid object and every once
in a while, say, in a few steps, it “describes” its “line of sight
(LoS)” in a convenient way . It can eventually perform a
detailed (high-resolution) description since it is quite small
compared to the object. So cumulating all the intermediate
LoS descriptions in a (2-D) histogram, particularly focusing
on continuous branches and major corners, yields an efficient
cue about the shape. Such a description is still feasible if some
portion of the object boundary is missing and this is
essentially the major advantage of this method. The
description frequency (i.e., how often the ant makes a new
intermediate description) and the length of LoS will obviously
be the two major parameters of this scheme.
(Impact Factor: 0.997)
809
International Journal of Emerging Technology & Research
(www.ijetr.org) ISSN (E): 2347-5900 ISSN (P): 2347-6079
Volume 1, Issue 4, May-June, 2014
The third one is the amount (number) of relevant sub segments
that are taken into consideration (description). Keeping this
number sufficiently low yields the method to describe only the
major object boundaries whilst discarding the texture edges.
Alternatively keeping this number high enough will allow the
proposed method to perform as a texture descriptor.
Retrieval of images is complicated by a lack of knowledge of
how people search for, and use, images. As the number of
images available increases, the more difficult it becomes to
find the image that meets a specific information need. In
addition, many of the documents that are being converted into
electronic formats contain images. Traditional retrieval and
indexing methods for providing access to large text databases
do not offer adequate access to the images. Similarity of
pictures and objects in pictures is reviewed for each of the
feature types, in close connection to the types and means of
feedback the user of the systems is capable of giving by
interaction.
2. Proposed Method
The proposed method is fully automatic (i.e., without any
supervision, feedback or training involved). Forming the
whole process as a Feature extraction (FeX) module into
MUVIS framework allows to test the overall performance in
the context of multimedia indexing and retrieval.
2.1 Related Work
Gabor filter is a widely used feature extraction method,
especially in image texture analysis for medical. The selection
of optimal filter parameters is usually problematic and unclear.
This study analyses the filter design essentials and proposes
two different methods to segment the Gabor filtered multichannel images. Another approach, the so-called angular
radial partitioning (ARP) . ARP basically works over radial
blocks . Although rotation invariance can be obtained with this
method, the shape outlines are degraded due to the loss of
aspect ratio during rescaling of the image into square
dimensions to fit a surrounding circle.
2.2 Automated Process
To extract (most) relevant subsegments over which the 2-D
WAH description is applied, a pre-processing phase is
performed, which mainly consists of four major parts: Frame
resampling, bilateral filtering and scale-map formation over
Canny edge fields, sub segment formation and analysis, and,
finally the selection of the relevant subsegments using a
relevance model. It makes the retrieval process more simple
and efficient so that it reduces the overload of the Multimedia
Database to provide a reliable process. The all process of this
paper has been shown in Fig:2.The first and natural step is
resampling into a practical dimension range, which is
sufficiently big for perceivable shapes but small enough not to
cause infeasible analysis time. The resampled frame can then
be used for multiscale analysis to form the scale-map, which is
entirely based on the adaptive canny edge detection over
nonlinear bilateral filter.
MPEG-7 edge histogram (EHD), generates a histogram of the
main edge directions (vertical, horizontal and two diagonals)
within fixed size blocks. It is an efficient texture descriptor for
the images with heavy textural presence. It can also work as a
shape descriptor as long as the edge field contains the true
object boundaries and is not saturated by the background
texture . In this case, the method is particularly efficient for
describing geometric objects due to its block-based edge
representation only with four directions.
The shape of an object is extracted properly and semantically
intact, several descriptors, which can be built from Fourier
transform, Hough transform , wavelet transform , curvature
scale space, Zernike moments , etc., can conveniently be
extracted either over the shape boundaries or the entire area
(the region of the object shape). Most of these methods
achieve a significant performance in terms of retrieval
efficiency and accuracy in binary shape databases; however,
especially in large multimedia databases containing ordinary
images or video clips, extraction of the true shape information
from natural objects first requires an automatic and highly
accurate segmentation, which is still an open and ill-posed
problem because the semantic objects in natural images do not
usually correspond to homogenous spatial regions in colour or
texture.
© Copyright reserved by IJETR
(Impact Factor: 0.997)
Figure 2 Overview of Automated Process
810
International Journal of Emerging Technology & Research
(www.ijetr.org) ISSN (E): 2347-5900 ISSN (P): 2347-6079
Volume 1, Issue 4, May-June, 2014
Bilateral filtering produces no phantom colors along edges in
color images, and reduces phantom colors where they appear in
the original image. Bilateral filter is a nonlinear filter which
depends on the image values. It is especially efficient in
removing the details when used iteratively, since the strong
edges remain with good localization. It smoothes the image
whilst preserving the edges.
3.1 Algorithm used to find for Corner Detection
The Algorithm used to find for the corner detection consists of
two steps. First step deals with the Extraction of Potential
Corners. While the next step deals with the Extraction of True
Corners.
Extraction of Potential Corners: Corner detection is
(1)
where, Iout is Output Image and Iin is Input Image, define the
domain standard deviation values ,σr defines the range
standard deviation values.
Canny edge detection is applied for each scale (after bilateral
filtering) ; however, its low-pass filter (Gaussian) is only
applied once to the input image in order not to cause excessive
blurring on the higher scales. Therefore, both (edge)
localization and scale information can be obtained from the
scale-map. The weight and length of a particular subsegment
signifies its relevance and, therefore, can conveniently be used
in the relevance model.
The relevant sub segment, which bear major object edges, are
usually longer with higher scale weights. On the contrary, the
irrelevant ones, which belong to details (texture, noise,
illumination variations, etc.), are usually shorter with lower
scale weights, since bilateral filter is likely to remove them in
the early iterations. Therefore, the relevancy, R, of a sub
segment SS, can then be expressed as follows:
R(SS)=W(SS) x L(SS)
(2)
where, W(SS) is the scale weight and L(SS) is the length
(total number of edge pixels) of SS. The next process is the
formation of 2-d walking ant histogram in Closed Loop or
Non Closed Loop form and all in one-pixel thick are used for
the formation of 2-D WAH, which is the union of two 2-D
histograms, each with equal dimensions. It is necessary to find
the corners and branches of the relevance model.
performed during the ant’s walk over each subsegment. The
bending ratio (BR) is calculated within the section. Each
section is traced over a subsegment from one end-point to the
other, and at each step, the bending ratio (BR) is calculated
within the section.
BR(p1)=LS/d∞(p1,p2)
(3)
where, p1 and p2 be the first and the last pixels to be
examined for a corner presence. LS be the pixel value. d∞
represents the distance in L∞ norm.
Extraction of True Corners: Basically, in Step 1, all
(potential) corners yielding a peak in BR plot are detected.
Step 2 is an optional step, which can post process them and
choose only the major corners among the ones that are too
close for visual perception. Therefore, apply non maximum
suppression in order to favour the one with highest corner
factor, which is the dot product of the bending ratio and the
curvature value. Let k(PiC) the corner factor of the CF(P iC)
potential corner
, be expressed as follows:
CF(PiC)=BR(PiC).k(PiC)
(4)
If there are n corners detected in a close vicinity where,
the
corner with highest is kept whilst the others are suppressed. In
order to accomplish this, let be the ith corner in between two
neighbor corners, pi-1C and pi+1C. Then CW(PiC) the corner
weight for the ith corner can thus be defined as follows
CW(PiC)=min[Np(PiC.Pi-1C) Np(PiC.Pi+1C)]/LSS
(5)
TCW= Σ PiCϵ SS CW(PiC)≤1
(6)
3. Corner Detection
where, Np(PxC, PyC) is the number of pixels between two
corner center pixels, PxC and PyC and LSS is the total number of
pixels in the subsegment, that is under 2-D WAH extraction
process. TCW represents the total corner weight, which
basically represents the corner amount of subsegment.
Corners can be defined as interest points where a radical
change occurs in the direction of shape boundary. Corners can
thus be found in locations where a discontinuity occurs in the
direction of a smooth section. Furthermore, it should eliminate
or minimize all false corners and be robust against noise, and
invariant to resolution, scale and orientation.
Once all corners are detected along with their weights, both
LoS sections (in forward and reverse directions) of each
corner are represented on the corner WAH bins with their
particular corner weights. Therefore, major corners will
eventually have dominance in the histogram compared to
minor ones, as intended.
© Copyright reserved by IJETR
(Impact Factor: 0.997)
811
International Journal of Emerging Technology & Research
Volume 1, Issue 4, May-June, 2014
(www.ijetr.org) ISSN (E): 2347-5900 ISSN (P): 2347-6079
4. Feature Extraction
4.4 Query by Example (QBE)
In order to test the retrieval efficiency of the proposed
descriptor over multimedia databases, the algorithm used into
a dynamic FeX module to be used for indexing and retrieval
processes in MUVIS framework. MUVIS [15] is a generic
framework over which any FeX method can be implemented
independent from the core of the system.
Query by example is a query technique that involves
providing the CBIR system with an example image that it will
then base its search upon. The underlying search algorithms
may vary depending on the application, but result images
should all share common elements with the provided example.
4.5 Query by Sketch (QBS)
4.1 Normalization of automated process
Once all corner and branch 2-D WAHs are extracted from the
relevant subsegments (both CL and NCL), they become
subject to a (unit) normalization process.
4.2 Sketch Similarity-Based Retrieval
The retrieval process in MUVIS is based on the traditional
query by example (QBE) scheme. The features of the query
item are used for (dis-) similarity measurement among all the
features of the (visual) items in the database.
D(q,x) = αc Dc (q,x) + (1-αc )DB (q,x)
The total dis- similarity distance D(q,x), calculated from its
branch, DB (q,x) and its corner, DC (q,x) components can be
expressed as in (9), where αc is the weight for corner
histogram differences in D(q,x) calculation. One can set α c ᷉
Avg (TCWq,TCWx) whenever accurate CL segmentation is
possible (e.g., for queries in a binary shape or natural image
database where the CL segments for objects are already
extracted as in [9]) since the corner information is complete
and reliable, otherwise it should be set to a small empirical
value (i.e., αc<0.25) in order to give more weight to branches
since some major corners might be missing on the NCL
subsegments present.
4.3 Content-based image retrieval (CBIR)
Content-based image retrieval (CBIR), also known as
query by image content (QBIC) and content-based visual
information retrieval (CBVIR) is the application of computer
vision to the image retrieval problem, that is, the problem of
searching for digital images in large databases. The storage of
image data is relatively straightforward, but accessing and
searching image databases is intrinsically harder than their
textual counterparts. The goal of Content-Based Image
Retrieval (CBIR) systems is to operate on collections of
images and, in response to visual queries, extract relevant
image. The application potential of CBIR for fast and effective
image retrieval is enormous, expanding the use of computer
technology to a management tool. Content-based means that
the search objects will analyse the actual contents of the
image.
© Copyright reserved by IJETR
The idea is to proceed interactively with the database system,
letting the system help the user in making an appropriate
sketch. The user begins by sketching an easy shape. This
approach compares the amount of edge pixels in the partitioned
black and white sketch and the target images.
The application of the homogeneous images, image regions
and an adapted distance measure is used, to overcome this
weakness. With the usage of the Average Normalized Modified
Retrieval Rank (ANMRR) this improvements are evaluated.
5. Experimental Results
The first experiments are performed to evaluate the accuracy
of the corner detector with respect to subjective test. The
retrieval performance of the proposed FeX module via QBE
scheme within a set of image databases is examined. Both
shape and texture based retrievals are evaluated using the
ground-truth methodology whilst providing both visual and
numerical results.
5.1 Sample databases
In this Experiment we have used three sample databases which
are based on Corel, Shape and Texture.
Corel_20K Image Database: There are 20000 images from
Corel database bearing similar content as in Corel_10K.
Shape Image Database: There are 1400 black and white
(binary) images that mainly represent the shapes of several
objects such as animals, cars, accessories, geometric objects.
Texture Image Database: There are 1760 texture images
that are obtained from brodatz database
Table 1. Computation Of Normalization Values
Images
2D WAH
Gabor
Rank 1
30.4192
26.7130
Rank 2
89.2445
85.2244
Rank 3
152.0397
148.3696
Rank 4
225.6708
225.4636
Rank 5
250.1665
250.2893
(Impact Factor: 0.997)
812
International Journal of Emerging Technology & Research
Volume 1, Issue 4, May-June, 2014
(www.ijetr.org) ISSN (E): 2347-5900 ISSN (P): 2347-6079
process; however, this is not a major problem since the feature
extraction is applied only once during the database indexing.
6. Conclusion
The successful results of this paper is implemented with a new
technology for edge detection based on multiscale subsegment
analysis over (Canny) edge field and, therefore, it can
conveniently work over arbitrary images, which may
encapsulate one or more objects in an inhomogeneous
background possibly with a strong textural structure. Use of
multimedia database would provide many improvements to
the current system of medical record keeping.
Figure 3: Medical Database
The Shape database is mainly used to examine the efficiency
and accuracy of the 2-D WAH’s shape descriptor whenever
CL segmentation is feasible. Finally, the retrieval evaluation is
presented over Texture database provided that 7. References
The 2-D WAH is tuned as a texture descriptor. The Corel
database with different sizes is to test the generality and
scaling capability of the proposed shape descriptor with
respect to the (increasing) database size. The graph is
generated based on the Rank of the Gabor and 2D WAH
values. Based on three different images, the time taken for
retrieval of images has been generated and the graph has
drawn.
Table II. Retrieval Time Of Different Images Using Different Methods
IMAGES
CBIR
QBE
QBS
Shape
0.98571
0.96613
0.93981
Texture
1.02872
0.98989
0.95869
Natural
1.17743
1.05781
0.98349
The proposed method basically achieves that the overall
algorithm is unsupervised and it is a fully automatic process.
This paper showed theoretically and by experiments that 2-D
WAH, it deals directly with arbitrary (natural) images without
any unreliable segmentation or object extraction preprocessing
stage. It has a simple, yet efficient, corner detector, which
improves the description power especially over CL segments
and also in sketched image, hence, it achieves generality and
robustness.
References
[1] M. Abdel-Mottaleb, “Image retrieval based on edge
representation,” in Proc. Int. Conf. Image Processing,
Piscatway, NJ, 2000, vol. 3, pp. 734–737.
[2] M. Bober, “MPEG-7 visual shape descriptors,” IEEE
Trans. Circuits Syst. Video Technol., vol. 11, no. 6,
pp. 716–719, Jun. 2001.
[3] Chalechale and A. Mertins, “An abstract image
representation based on edge pixel neighbor-hood
information (EPNI),” Lecture Notes Comput. Sci.,
vol. 2510, pp. 67–74, 2002.
Forming the whole process as a Feature extraction (FeX)
module into MUVIS framework allows to test the overall
performance in the context of multimedia indexing and
retrieval from Shape, Corels and Texture databases. Since
number of scales can vary, per-scale time for 2-D WAH
feature extraction for Corels.
Furthermore, no corner detection is applied for Texture
database; therefore, the fastest time is achieved for this
database, whereas the rest of the feature extraction process
only takes a fraction of a second. Compared to other
competing methods, 2-D WAH usually has the slowest
indexing time, mostly due to nonlinear Bilateral Filtering
© Copyright reserved by IJETR
(Impact Factor: 0.997)
813