Integrated Intelligent Research (IIR) International Journal of Business Intelligents

Volume: 05 Issue: 01 June 2016 Page No.96-100

ISSN: 2278-2400

Feature Extraction in Content based Image Retrieval

R. Ramya 1,C.Kalaiselvan2
PG Scholar, Dept. Of ECE, Pavendar Bharathidasan College of engineering and technology,Trichy, Tamilnadu, India 1
Assistant Prof, Dept. of ECE, Pavendar Bharathidasan College of engineering and technology,Trichy, Tamilnadu, India 2
Email:ramyabehira@gmail.com,kalaiaswath@gmail.com

Abstract-A technique for Content Based Image Retrieval recorded in the storage in the compressed form. Without
(CBIR) for the generation of image content descriptor which performing the decoding process the feature extractor simply
exploiting the advantage of low complexity Order Dither Block generate an image feature for CBIR task. BTC is also an image
Truncation Coding (ODBTC). The quantizer and bitmap image compression method which requires simple process in both the
are the compressed form of image obtained from the ODBTC encoding and decoding stages. The important and first step in
technique in encoding step. Decoding is not performed in this BTC is divide the image into several block. That block is
method. It has two image feature such as Color Co-occurrence represented with the two specific quantizer which maintain it
Feature (CCF) and Bit Pattern Feature (BPF) for indexing the mean value and standard deviation to original image block. The
image. These features are directly obtained from ODBTC BTC produces two quantizer such as high and low quantizer.
encoded data stream. By comparing with the BTC image The bitmap image are generated at end of decoding process. The
retrieval system and other earlier method the experimental result BTC replacing the bitmap information with high and low
show the proposed method is superior. ODBTC is suited for quantizer which is reverse procedure. In encoding and decoding
image compression and it is a simple and effective descriptor to stage the BTC never requires the codebook information in VQ
index the image in CBIR system. Content-based image retrieval image compression or the quantization table in JPEG which are
is a technique which is used to extract the images on the basis of auxiliary information. The size of DataStream is required using
their content such as texture, color, shape and spatial layout. In entropy coding in the BTC method which maintain acceptable
order to minimize this gap many concepts was introduced. visual image. The first CBIR system developed using BTC is
Moreover, Images can be stored and extracted based on various explained in the color image indexing using BTC method. This
features and one of the prominent feature is Texture. method explain the nature of BTC to generate the image feature
using two quantized value and bitmap image. In the early work
to index a set of image in the database two image feature have
Keywords: Content Based Image Retrieval, Neural Network,
been proposed such as block color co-occurrence matrix and
Relevance Feedback, Color, Texture, Shape, Semantics.
block pattern histogram. This method also utilize the RGB color
space and YCbCr color space. This method having a better
I. INTRODUCTION result in term of retrieval accuracy compared to the former
An image retrieval system returns a set of image from the method. The RGB color space is used for extraction of image
database. Database having the collection of image. The image feature descriptor. In the BTC encoding is performed on each
retrieval is to meet users demand which having similarity color space such as red, green and blue separately.
evaluation such as image content similarity, edge pattern
similarity, color similarity, etc. In real time application an image The color channel is used to extract the bit pattern codebook and
retrieval system is an efficient way to access, browse and traditional histogram. In BTC indexing method a different
retrieve a set of similar image. In the earlier method the image is approach for CBIR system using color moment and k-mean
simply computed in the DCT domain.JPEG standard clustering can be found. In the CBIR application the BTC
compression is an improvement of image retrieval in DCT scheme is the convenient and effective method to index image in
domain. Most of image are dealing with MPEG-7 visual content database. In image coding BTC had played an important role.
descriptor including the Color Descriptor (CD), Texture Many advanced coding technique is inspired by BTC for its
Descriptor (TD) and Shape Descriptor (SD) to establish the stability and simplicity. To reduce the computational
standard for CBIR task. This standard provide a great advantage complexity, improve image quality and higher compression ratio
in CBIR research field and also having benefit in which image many improvement are made in BTC scheme. HBTC and BTC
content descriptor is modified by user. The image descriptor is having computational simplicity which made it is an attractive
required for modification and Recalculation so that the original tool in the real time implementation. HBTC is derived from
image is not necessarily transferred. For recognizing the same BTC in which the bitmap image is replaced by halftone image.
categories of image from the huge database a new type of CBIR The main different between this technique is it quantizer
approach is presented in the spatial pyramid and order less bag determination. The BTC maintain the mean value and standard
of feature image representation. This is mainly used for natural deviation but HBTC quantizer is obtain from minimum and
scene classification. Receptive field learning for pooled image maximum value in each image block. An example of HBTC is
feature for image classification. This having much lower feature dither based BTC in which bitmap is generated from dithering
dimensionality compared to the other former scheme. The image approach. The dithering based BTC is the Order Dither Block
feature descriptor is extract from compressed data stream in Truncation Coding (ODBTC) which involves the HVS for
CBIR system. For reducing the storage space the image are achieving image quality. The ODBTC scheme is uses the dither
96
Integrated Intelligent Research (IIR) International Journal of Business Intelligents
Volume: 05 Issue: 01 June 2016 Page No.96-100
ISSN: 2278-2400
array look up table in encoding stage. The extreme value in the compression method BTC-VQ is used. The new method for
ODBTC are obtained from maximum and minimum value in the image retrieval is implemented based on color histogram and
image block. Water making scheme are also used in this vlock pattern histogram.
technique which is used for application requiring privacy and
The key step for locating the embedded information bit they
ownership protection. From the ODBTC compressed data
uses inverse half toning and second round of halftonning. The
stream a new approach is proposed to index an image in
objective good quality image with flexible capacity and
database. This method of image retrieval system generate two
reasonable complexity is obtained from the experimental
image feature such as CCF and BPF from the color quantizer
result. The correct decoding rate of 100% have maintained in
and bitmap image.The accumulation of these image collections
this technique. The decoder reconstruct the original hotspot
(including art works, satellite and medical imagery) is attracting
ODBTC image which boost the flexibility in the image
more and more users in various professional fields for example
control. It have the original image unmodified. BTC offering
geography, medicine, architecture, advertising, design, fashion
good image quality by the efficient compression method. The
and publishing. Image retrieval is concerned with techniques for
high compression ratio application uses BTC scheme which
storing and retrieving images both efficiently and effectively.
causes severe perceptual atrifacts of blocking effect. LUT is
Early image retrieval methods locate the desired images by
used to reduce the complexity of BTC. For effectively
matching keywords that are assigned to each image manually.
displaying of ODBTC image some new progressing scheme is
However, as a result of the large number of images in
used. To determine the transmitting order ODBTC utilize the
collections, manual processing has become impractical.
bit interleaved bitmap image technique which reconstruct the
Content-Based Image Retrieval (CBIR) offers a convenient way
ODBTC image. From the ancient times, the use of the images
to browse and search the desired image in the huge image
is very common. They tells us who we are and where we
database. The CBIR employs the image features of visual
come from. Images are the best way to memorize the time that
content to represent and index the image in database. These
we spent with our loved ones. Also, images are more eye-
features can be color, texture, shape, etc. The feature choice
catchy then the raw data. Thus now a day’s images play a very
depends on the user’s preference or is decided by the expert-
crucial role in every field. With the advancement in
system. Finding a single best representative feature of an image
technology, everything becomes digitized which provides ease
is very difficult because of the fact that the photographer may
to the user in every aspects. World Wide Web is the real
take several images under different conditions such as different
example for the enormous storage of the digitized images. It
lighting sources, various view angles, different illumination
contains millions or trillions of images in their databases.
changes, etc. Developing an effective and efficient image feature
Digitization not only makes the process to store the images in
descriptor becomes a challenging task for CBIR system to
an easy manner but also provide you to search it an efficient
achieve a high image retrieval performance. Many attempts and
manner. Many researchers have found that visual data all over
researches have been devoted to improve the retrieval accuracy
the world is increasing day by day at very fast speed.
in the CBIR system. One of these efforts is employing an image
feature descriptor derived
III. CBIR APPROACH
II. RELATIVE WORKS
Content Based Image Retrieval (CBIR) is the method of
Embedding watermarks into compressed image a new method retrieving images from the large image databases as per the
which is based on BTC and halftonning technique is user demand. It is also known as Query by Image Content
introduced. The very simple and efficient image compression (QBIC) and Content Visual Information Retrieval (CBVIR).
technique is BTC. When the image block size is increases In CBIR, content based means the searching of image is
because it produce image of high quality and with high proceed on the actual content of image rather than its
blocking effects. With the same compression capability a new metadata. The Content Based Image Retrieval System is used
method is introduced to solve the problem is ODBTC. Embed to extract the features, indexing those features using
the watermark into compressed image is also done by this appropriate structures and efficiently provide answers to the
technique. Order dithering is used to carry out watermarking user’s query. To provide the satisfactory answer to the user
also to incorporate void-and-cluster method. The other query, CBIR provides some flow of work. Firstly CBIR
method based on BTC and VQ has an effective feature for system takes the RGB image as an input, performs feature
color image.in this technique input color image is separated
extraction, performs some similarity computations with the
into Y, Cb and Cr components. BTC is performed by 4x4 Y
block. This produces mean pair sequence and bit plane images stored in database and retrieves the output image on
sequence. To obtain the contrast and visual pattern co- the basis of similarity computation. There are some basic
occurrence matrix they are quantified by contrast pattern CBIR fundamentals and are divided into three parts such as
codebook and visual pattern code book. VQ is performed by feature extraction, multidimensional indexing and Retrieval
4x4 Cb blocks and Cr block respectively. This produces color system architecture.A new type of CBIR approach is
pattern co-occurrence matrix. This having high retrieval presented in which the spatial pyramid and order less bag-of
performance. The database consist of 9983 image with features image representation were employed for recognizing
different content is used to examine in this process and it the scene categories of images from a huge database. This
result have been compared with similar method. For
97
Integrated Intelligent Research (IIR) International Journal of Business Intelligents
Volume: 05 Issue: 01 June 2016 Page No.96-100
ISSN: 2278-2400
method offers a promising result and outperforms the former quantization is used to generate bit pattern codebook and
existing methods in terms of the natural scene classification. many bitmap images are involved in the training stage.in the
The method presented the holistic representation of spatial code book generation all the all the code vector have value
envelop with a very low dimensionality for representing the between zero(black pixel) and one(white pixel) as opposed to
scene image. This approach presented an outstanding result in binary value. The hard thresholding performs the binarization
the scene categorization. The method in proposed a new of all code vector for final result. From the similarity
approach for image classification with the receptive field measurement between the bitmap and code word is used to
design and the concept of over-completeness methodology to generate bitmap of each block. The feature dimensionality of
achieve a preferable result. As reported this method achieved the bit pattern feature is always identical to the bit pattern
the best classification performance with much lower feature code book sizes. The overall dimensionality is based on the
dimensionality compared to that of the former schemes in feature descriptor. The CCF and BPF having the similar
image classification task.The CBIR system which extracts an computation making it always suitable for faster response of
image feature descriptor from the compressed data stream has the real time applications.
become an important issue. Since most of the images are
recorded in the storage device in compressed format for IV. PROPOSED SYSTEM
reducing the storage space requirement.The Block Truncation
Coding (BTC) is an image compression method which FEATURE EXTRACTION:
requires simple process on both encoding and decoding Features are divided into two categories respectively text
stages. The BTC compresses an image in a simple and based and visual based. Textual features are keywords, tags,
efficient way. BTC firstly divides an input image into several annotations etc. Visual features are color, space and texture
etc. Visual features are the important features of an image for
image blocks, and each image block is subsequently
pattern recognition.
represented with two specific quantizes to maintain its mean
value and standard deviation identical to the original image
block.

The BTC produces two quantizers, namely high and low

quantizers, and a bitmap image at the end of the decoding
process. The BTC decoding performs the reverse procedure
by simply replacing the bitmap information with the high or
low quantizer. BTC never requires auxiliary information
during encoding and decoding stages such as the codebook
information in the Vector Quantization (VQ) image
compression or the quantization table in JPEG. The BTC
maintains acceptable visual image, and the size of the data
stream can be further reduced using the entropy coding.
The information about the image content is obtain from the COLOR BASED RETRIEVAL
color distribution of the pixels of image. To obtain the image
The Basic technique which is used is based on the technique
color distribution color co-occurrence matrix is used. The
of color histogram. Color Histogram of each image is
occurrence probability of pixel along with its adjacent
calculated and then stored in the database which represents the
neighbour is calculates from the colour co-occurrence matrix. proportion of pixel of each color within the image. Then
The spatial representation of image is also obtain from the matching algorithm will extract those images from the
matrix. CCF is computed from the ODBTC color databases whose color histogram matches with the required
quantizers .the specific code book is used to index maximum one. There are various types of histograms: normal, weighted,
and minimum quantizer. The color co-occurrence matrix is dominant, and fuzzy, various color spaces: HSV, grayscale,
obtained from the above indexed values. In RBG indexing HSL, Lab, Luv, HMMD, and YCbCr.
process the RGB pixel is mapped three tuples into finite
subsets. The color indexing process of ODBTC minimum TEXTURE BASED RETRIEVAL
quantizer nearly equal to the minimum quantizer of each
block. The color co-occurrence matrix is the sparse matrix. In This is a very important characteristic of an image because it
this zero dominates its entries. In order to increase the speed is able to distinguish two images with same color and shape.
Variety of techniques has been proposed for matching the
and to reduce the feature dimensionality of CCF the color co-
texture similarity. Tamura et al. proposed a texture
occurrence matrix is binned along its columns or rows to form
representation on 6 statistical features, including, coarseness,
a 1D image feature descriptor. The CCF calculation is simple contrast, directionality, line-likeness, regularity, and
and it is more preferable for CBIR task.Another feature which roughness. These features were considered to be the most
character the edge, shape and image content is bit pattern visually meaningful. Various techniques designed for texture
feature. The bit pattern code book is produced by binary feature extractions are: statistical parameters, entropy
vector quantisation from the set of training bit map images measures, transformed spaces and Markov Hidden Fields
obtained from the ODBTC encoding process. Binary vector algorithms.
98
Integrated Intelligent Research (IIR) International Journal of Business Intelligents
Volume: 05 Issue: 01 June 2016 Page No.96-100
ISSN: 2278-2400
based on diminutive waves called wavelet of varying
SHAPE BASED RETRIEVAL frequency and limited duration. Discrete Wavelet transform
divides the images into four different parts namely higher
This is a well-defined term, which refers to the shape of the frequency part (HH), High Low Frequency part (HL), Low
image. This is feature which naturally distinguishes the High Frequency part (LH), and Lower frequency part (LL).
images. There are two main features of the shape: Global After doing the vertical parts as 1-level images
feature (like aspect ratio) and local feature (like boundary decomposition, it computes moments of all parts and store and
segments). Shape of an image can be represented using area, use it as feature to obtain images.
perimeter, radiuses, skeleton, statistics moments, form
signature, Fourier and Hough contour signature. GABOR FILTER

V. SYSTEM DESIGN AND IMPLEMENTATION It is widely used for texture analysis because of its similar
Multidimensional indexing techniques are mainly used to characteristics with human perception. A two dimensional
make the CBIR truly scalable large size image collection. Gabor function g(x, y) consists of a sinusoidal plane wave of
Most of the images are having high dimensionality. So the some frequency and orientation (Carrier), and two
best way to index such images is to reduce the dimensionality dimensional translated. Gaussian Envelope is used to
and then indexing the images. For dimension reduction, modulate it.
clustering is used. Clustering can be used in various forms like
pattern recognition, speech analysis and information retrieval. SUPPORT VECTOR MACHINE
Clustering can be performed row wise as well as column wise
to perform recognition or grouping. Images are indexed after It is supervised learning technique in which data is analyzed
feature extraction and then similarity measurement is and identify pattern used for classification purpose. In
performed. Similarity evaluation is done between the features classification it takes set of input, read it and forms output for
of the query image and the features of the target image in the each desired input and if the output is continuous then
database. Similarity measure computes the degree of regression is performed.The color co-occurrence feature and
similarity between a pair of images. It represents the distance bit pattern feature having different modalities but they
between feature vectors representing the images. Similarity combining the features and also determining their similarity
images should have smaller distance between them and constant carried in the experiments. The target image is the
different images should have large distance.The relative scaled version of query image. Lot of experiments were
distance measure is used to measure the similarity between conducted to check the performance of this method. The
two image (query image and target image).The important role ODBTC encoded data stream produces image descriptor is
for receiving the set of similar image are played by the already stored in the database. The overall image in the
similarity distance. The query image is encoded by the database are computed by CCF and BPF. Based on the
process of ODBTC which produces corresponding color co- similarity distance the system returns a set of similar image
occurrence feature and bit pattern feature. Then the feature of from the database. When several image are turned as queries
target image is compared with the above features. A set of the image retrieval performance is tested. The effectiveness of
similar image to the query image were returned based on proposed method and former existing method are measured by
similarity distance score. The lowest score which indicate the performance evaluation. A set of similar image to the query
similar image to the query image. image were returned based on similarity distance score. Four
quantitative evaluation is used to find the performance such as
precision, recall, average retrieval rate and average normalised
modified retrieval rank. The performance is measured with
proportion correction classification from the nearest
neighbour classifier in the image classification task.as used in
the image retrieval task the classifier assign the class label
using similarity distance computation. In the database the
similarity distance is stored in the ascending order between
the query image and database image. The performance
evaluation is conducted by averaging averaging the values of
overall query image. All the images are turned into query
image in the image retrieval system. The average precision
and average recall measurement are used for describing the
image retrieval performance. The higher value of the
precision, recall and average retrieval rate denotes the higher
retrieval rate and better performance of system. Binary vector
quantization is used to generate bit pattern codebook and
SEMANTIC TEMPLATE: many bitmap images are involved in the training stage.in the
code book generation all the all the code vector have value
This technique is generated to support high level image between zero(black pixel) and one(white pixel) as opposed to
retrieval and not so widely used. This technique is usually binary value. The hard thresholding performs the binarization
defined as the representative feature of concept calculated of all code vector for final result. From the similarity
from a collection of sample images. Wavelet transform is measurement between the bitmap and code word is used to

99
 Integrated Intelligent Research (IIR) International Journal of Business Intelligents
Volume: 05 Issue: 01 June 2016 Page No.96-100
ISSN: 2278-2400
generate bitmap of each block. The feature dimensionality of [13] TK Rama Krishna Rao, R Usha Rani, “Content Based Image Retrieval
Through Feed Forward Neural Networks”, IEEE 8th International
the bit pattern feature is always identical to the bit pattern Colloquium on Signal Processing and its applications, 2012.
code book sizes. The overall dimensionality is based on the [14] Ionut Mironica and Radu Dogaru, “A comparison between various
feature descriptor.The bitmap image are generated at end of classification methods for image classification stage in CBIR”, 2008.
decoding process. The BTC replacing the bitmap information [15] Fazal-e-Malik, Baharum Baharudin, “Efficient Image Retrieval Based
on Texture Features”, IEEE, 2011.
with high and low quantizer which is reverse procedure. In [16] Wu Kai-xing, Xu Qiang, “Image retrieval based on fuzzy color
encoding and decoding stage the BTC never requires the histogram”, International Conference on Intelligent Information
codebook information in VQ image compression or the Hiding and Multimedia Signal Processing, IEEE, 2008.
quantization table in JPEG which are auxiliary information. [17] Mircea Ionescu, Anca Ralescu, “Fuzzy Hamming Distance in a
Content- Based Image Retrieval System”, 25-29 July, 2004, Budapest,
The size of DataStream is required using entropy coding in Hungary, IEEE.
the BTC method which maintain acceptable visual image. The [18] P.S. Hiremath, Jagadeesh Pujari, “Content Based Image Retrieval
first CBIR system developed using BTC is explained in the based on Color, Texture and Shape features using image and its
color image indexing using BTC method. complement”, International Journal of Computer Science and Security,
Volume(1) : Issue(4).
[19] Timo Ojala et.al, “CMRS: Architecture for Content-Based Multimedia
VI. CONCLUSION Retrieval”.
[20] Myron Flickner et.al, “Query by image and video content: The QBIC
System”.
As the Content Based Image Retrieval technique is applied
only on images, in proposed system this CBIR technique is
applied on videos. Video is a collection of image frames so
each individual frame is considered as the single image.
Feature extraction is done on each image frame. This
approach is used only for the live streaming video of .avi file
format. In this approach the ODBTC indexing technique is
used for indexing the images but in video, the image frames
are already in sequence so ODBTC is directly applied on the
predefined sequence of image frames in video.
REFERENCE
[1] Michele Saad, Content-Based Image Retrieval – A Literature Survey,
2008.
[2] Datta, R., Li, J., and Wang, J., Content-Based Image Retrieval – A
Survey on the approaches and trends of the New Age, Proceedings of
ACM International Retrieval, ACM Multimedia, Singapore, pp. 77-82,
2005.
[3] S. Kulkarni and B. Verma, “Fuzzy logic based Texture Queries for
CBIR”, IEEE proceedings of the fifth International Conference on
Computational Intelligence and Multimedia Applications
(ICCIMA’03), 2003.
[4] Hsin-Chih Lin, Chih-Yi Chiu, Shi-Nine Yang, “Finding textures by
textual descriptions, visual examples, and relevance feedbacks”,
Pattern Recognition Letters,24(12):2255-2267, October 2003.
[5] Heba Aboulmagd Ahmed, Neamat El Gayar, Hoda Onsi, “A New
Approach in Content- Based Image Retrieval using Fuzzy Logic”,
INFOS2008, march 27-29, 2008, Cairo-Egypt.
[6] William B. Thompson, “Textural Boundary Analysis”, IEEE
Transactions on computers, March 1977.
[7] Samuel Barrett, Ran Chang and Xiaojun Qi, “A Fuzzy Combined
Learning Approach to Content-Based Image Retrieval”, IEEE, 2009.
[8] Swarup Medasani and Raghu Krishnapuram, “A Fuzzy Approach to
Content-Based Image Retrieval”, IEEE International Fuzzy Systems
Conference Proceedings, August 22-25, 2009, Seoul, Korea.
[9] Dr. B. Prabhakara Rao et.al, “CTDCIRS: Content based Image
retrieval system based on Dominant color and texture features”, IEEE
International Journal of Computer Applications (0975 – 8887),
Volume 18 – No.6, March 2011.
[10] Wang Xiaoling et.al, “Application of the fuzzy logic in the Content-
based Image Retrieval”, Journal of Computer Science & Technology
Vol 5 No.1. April 2005.
[11] Rushikesh Borse et.al, “Efficient Implementation of CBIR System and
Framework of Fuzzy Semantics”, IEEE International Conference on
Advances in Mobile Network, Communication and its Applications,
2012.
[12] Hideyuki Tamura et.al, “Textural Features Corresponding to Visual
Perception”, IEEE Transactions on Systems, man, and cybernetics,
Vol. SMC – 8, No.6, June 1978.

100
Integrated Intelligent Research (IIR) International Journal of Business Intelligents
Volume: 05 Issue: 01 June 2016 Page No.96-100
ISSN: 2278-2400

101