Académique Documents
Professionnel Documents
Culture Documents
INTRODUCTION
Web-scale image search engines mostly use keywords as queries and rely on
surrounding text to search images. It is well known that they suffer from the ambiguity of
query keywords. For example, using apple as query, the retrieved images belong to
different categories, such as red apple, apple logo, and apple laptop. Online image re
ranking has been shown to be an effective way to improve the image search results . Major
internet image search engines have since adopted the re-ranking strategy. Its diagram is
shown in Figure 1. Given a query keyword input by a user, according to a stored word-image
index file, a pool of images relevant to the query keyword are retrieved by the search engine.
By asking a user to select a query image, which reflects the users search intention, from the
pool, the remaining images in the pool are re-ranked based on their visual similarities with
the query image. The visual features of images are pre-computed offline and stored by the
search engine. The main online computational cost of image re-ranking is on comparing
visual features. In order to achieve high efficiency, the visual feature vectors need to be short
and their matching needs to be fast.Another major challenge is that the similarities of low
level visual features may not well correlate with images high-level semantic meanings
which interpret users search intention. To narrow down this semantic gap, for offline image
recognition and retrieval, there have been a number of studies to map visual features to a set
of predefined concepts or attributes as semantic signature However, these approaches are
only applicable to closed image sets of relatively small sizes. They are not suitable for online
web- based image re-ranking. According to our empirical study, images retrieved by 120
query keywords alone include more than 1500 concepts. Therefore, it is difficult and
inefficient to design a huge concept dictionary to characterize highly diverse web images.
CHAPTER 2
LITERATURE SURVEY
2.1 IMAGE RETRIEVAL USING SEMANTICS OF QUERY IMAGE (2013)
/(SIDDANAGOWDA G. R)
The fields of image-based content retrieval and automatic image annotation are becoming
more and more relevant to the ways in which large libraries of digital media are stored and
accessed. As multimedia and imaging technologies improve, so does the wealth of visual data
in libraries, necessitating an automated mechanism with which to index and thus access
content by a variety of means Content based image retrieval for general-purpose image
databases is a highly challenging problem because of the large size of the database, the
difficulty of understanding images, both by people and computers, the difficulty of
formulating a query, and the issue of evaluating results properly. A number of generalpurpose image search engines have been developed The common ground for CBIR systems is
to extract a signature for every image based on its pixel values and to define a rule for
comparing images. The signature serves as an image representation in the view of a CBIR
system. Content Based Image Retrieval (CBIR) is any technology that in principle helps to
organize digital image archives by their visual content. By this definition, anything ranging
from an image similarity function to a robust image annotation engine falls under the purview
of CBIR the most common form of CBIR is an image search based on visual Features
Re-ranking with user interactions or active re-ranking is introduced in this system. Collecting
information from users to obtain the specified semantic space. Localizing the visual
characteristics of the users intention in this specific semantic space.
Limitations in this paper is that most of the image retrieval systems present today are textbased, in which images are manually annotated by text-based keywords and when we query
by a keyword, instead of looking into the contents of the image, this system matches the
query to the keywords present in the database. Re-ranking with user interactions Collecting
information from users to obtain the specified semantic space. Localizing the visual
characteristics of the users intention in this specific semantic space .These disadvantages of
text-based image retrieval techniques call for another relatively new technique known as
Content-Based Image Retrieval (CBIR). CBIR is a technology that in principle helps
2
organize digital image archives according to their visual content. This system distinguishes
the different regions present in an image based on their similarity in color, pattern, texture,
shape, etc. and decides the similarity between two images by reckoning the closeness of these
different regions.
images with similar content, most current algorithms fail to accurately relate its high level
concept, or the semantics of the image, to its lower level content. The problem with these
algorithms is their reliance on visual similarity in judging semantic similarity. Moreover,
semantic similarity is a highly subjective measure.Shape is a key attribute of segmented
image regions, and its efficient and robust representation plays an importantrole in retrieval.
A shape similarity measure using discrete curve evolution to simplify contours, is discussed
in [59].Doing this contour simplification helps to remove noisy and irrelevant shape features
from consideration.
2.3 INTENT SEARCH: CAPTURING USER INTENTION FOR ONE-CLICK
INTERNET IMAGE SEARCH.(YEAR 2014) /(XIAOOU TANG)
Many commercial Internet scale image search engines use only keywords as queries.Users
type querykeywords in the hope of finding a certain type of images.The search engine returns
thousands of images ranked bythe keywords extracted from the surrounding text. It is
wellknown that text-based image search suffers from theambiguity of query keywords. The
keywords provided byusers tend to be short.Adaptive Similarity is proposed, motivated by
the ideathat a user always has specific intention whensubmitting a query image. For example,
if the usersubmits a picture with a big face in the middle, mostprobably he/she wants images
with similar faces andusing face-related features is more appropriate. In ourapproach, the
query image is first categorized intoone of the predefined adaptive weight categories,such as
portrait and scenery.Keyword expansion. Query keywords input by user stand to be short
and some important keywords maybe missed because of users lack of knowledge on the
textual description of target images. In our approach ,query keywords are expanded to
capture users search intention, inferred from the visual content of query images, which are
not considered in traditional keyword expansion approaches.The image pool retrieved by
text-based search accommodates images with a large variety of semantic meanings and the
number of images related to the query image is small. In this case, re-ranking images in the
pool is not very effective. Thus, more accurate query by keywords is needed to narrow the
intention and retrieve more relevant images.Visual query expansion. One query image is not
diverse enough to capture the users intention. In Step 2, a cluster of images all containing the
sameexpanded keywords and visually similar to thequery image are found.All four of these
steps are automatic with only one clickin the first step withoutincreasing users burden. This
makes it possible for Internet scale image search by both textual and visual content with a
very simple user interface.
4
The drawback is that given a query image, similarity retrieval involves searching the database
for similar color distributions as the input query. Since the number of representative colors is
small, one can first search the database for each of the representative colors separately, and
then combine the results. Calculating Euclidian distance in a color space can do searching for
individual colors very efficiently. Each pixel is associated to a specific histogram bin only on
the basis of its own color, and color similarity across different bins or color dissimilarity in
the same bin are not taken into account.Since any pixel in the image can be described by
three components in a certain color space.
2.4 DESCRIPTOR LEARNING FOR EFFICIENT RETRIEVAL (JAMES PHILBIN) /
(2010)
There have been several recent applications of distance learning to classification problems
however these methods assume clean, labelled data indicating pairs of points that belong to
the same class and pairs that belong to different classes. In our task, even when the same
object appears in two images, the images typically have different backgrounds and there is
anon-trivial transformation between the views of a common object, so we cannot simply
classify images as being matching or non-matching. At the same time the number of
individual descriptors per image and the complexity of the correspondence problem between
them means that manually labelling the sets of matching and non-matching descriptors would
be unacceptably burdensome. Therefore, in this work, we introduce a new method for
generating training data from a corpus of unlabelled images using standard techniques from
multi-view geometry. In contrast to Hua et al. [16], who also generated training pairs from
unlabelled image data via patches matched by the Photo Tourism system here we adopt a
much cheaper pairwise image measure which doesnt require us to compute a global bundle
adjustment over many image pairs. Thus, we can train on patches of objects that appear in as
few as two images. Previous works in distance learning use two categories of point pairs for
training: matching and non-matching, typically derived from known class labels. In this
work, we show that we can significantly improve performance by forming two nonmatching categories: random pairs of features; and those which areeasily confused by a
baseline method. We adopt a margin-based cost function todistinguish these three categories
of points, and show that this gives improvedperformance more than using non-margin-based
methods To optimize this cost function, a fast, stochastic, online learning procedure is used
5
that permits the use of millions of training pairs. We will show that non-linear projection
methods, previously used for hand-written digit classification [13], perform better than the
linear projections previously applied to computer vision distance learning .The next section
motivates the distance learning task by showing that retrieval performance is significantly
worse using standard quantized descriptors than when a much slower, exhaustive search
procedure is applied to the raw SIFT descriptors this indicates the potential gain achievable
from better clustering. After describing in Section 3 how we automatically generate our
training data, we set out our learning methods in Section 4 and then conclude with results
andDescriptor Learning for Efficient Retrieval a discussion. Improved performance is
demonstrated over SIFT descriptors on standard datasets with learnt descriptors.In practice, it
is not possible to fully separate these pairwise distances because of noise in the training data
and restricted model complexity, so instead a margin based approach will be used which
encourages the distance between the three classes of point pairs to separate without enforcing
the distance ordering as a hard constraint. The loss function for this situation is The first
margin aims to separate the positive and NN negative point pairs confused by SIFT in the
original space. The second margin applies a force to the random negatives to keep them
distant from the positive pairs ideally the overlap in histograms between the positive and
random negative point pairs should be small.
Limitations in this paper is that image pair is chosen at random from the dataset. A set of
putative matches is computed between the image pair. Each putativematch consists of a pair
of elliptical features, one in each image, that pass Lowes second nearest neighbour ratio test
[18] on their SIFT descriptors; RANSAC is used to estimate an affine transform between the
images together with a number of inliers consistent with that transform. Point pairs are only
taken from image matches with greater than 20 verified inliers. The ratio test ensures that
putative matches are distinctive for that particular pair of images. This procedure generates
three sets of point pairs.These are the point pairs found as inliers by RANSAC. These are
pairs marked as outliers by RANSACthey are generally close in descriptor space as they
were found to be descriptor-space nearest neighbours between the two images, but are
spatially inconsistent with the best-fitting affine transformation found between the images.
Drawbacks is that the proposed re-ranking procedure where the dashed box represents the
Relevance Model Re-ranking box in the users input a query consisting of keywords to
describe the pictures they are looking for, and a web image search engine returns a rank list
of images. The same query is also fed into a web text search engine, and retrieved documents
are used to estimate the relevance model for the query Q. We then calculate the KL
divergence between the relevance model and the unigram mode of each document D
associated with the image I in the image rank list, and re-rank the list according to the
divergence. The re-ranking process based on relevance model still can improve the
performance, suggesting that global informationfrom the document can provide additional
cues to judge the relevance of the image. Internet users are usually with limit time and
patience, and high precision at top-ranked documents will save user a lot of efforts and help
them find relevant images not easily and quickly.
CHAPTER 3
SYSTEM ANALSIS
3.1 EXISTING SYSTEM
In this paper, a novel framework is proposed for web imagere-ranking. Instead of
constructing a universal conceptdictionary, it learns different visual semantic spaces for
differentquery keywords individually and automatically. Believe that the semantic space
related to the images to bere-ranked can be significantly narrowed down by the
querykeyword provided by the user. For example, if the querykeyword is apple, the
semantic concepts of mountainsand Paris are unlikely to be relevant and can be
ignored.Instead, the semantic concepts of computers and fruit will be used to learn the
visual semantic space related toapple. The query-specific visual semantic spaces canmore
accurately model the images to be re-ranked, sincethey have removed other potentially
unlimited number of non-relevant concepts, which serve only as noise and deteriorate the
performance of re-ranking in terms of both accuracy and computational cost. The visual
features of images are then projected into their related visual semantic spaces to get semantic
signatures. At the online stage, images are re-ranked by comparing their semantic signatures
obtained from the visual semantic space of the query keyword. Our experiments show that
the semantic space of a query keyword can be described by just 20
30 concepts (also
referred as reference classes in our paper). Therefore the semantic signatures are very short
and online image re-ranking becomes extremely efficient. Because of the large number of
keywords and the dynamic variations of the web, the visual semantic spaces of query
keywords need to be automatically learned. Instead of manually defined, under our
framework this is done through keyword expansions. Another contribution of the paper is to
introduce a large scale benchmark database1 with manually labeled ground truth for the
performance evaluation of image re-ranking. It includes 120; 000 labeled images of around
1500 categories (which are defined by semantic concepts) retrieved by the Bing Image
Search using 120 query keywords. Experiments on this benchmark database show that
20% 35% relative improvement has been achieved on re-ranking precisions with much
faster speed by our approach, compared with the state-of-the-art methods. Our experiments
show that the semantic space of a queryThis is the most common form of text search on the
Web. Most search engines do their text query and retrieval using keywords. The keywords
based searches theyusually provide results from blogs or other discussion boards. The user
9
cannot have a satisfaction with these results due to lack of trusts on blogs etc.low precision
and highrecall rate. In early search engine that offered disambiguation to search terms. User
intention identification plays an important role in the intelligent semantic search engine.
some method of
limited users effort to just one-click feedback. Such simple image re-ranking approach has
been adopted by popular web-scale image search engines such as Bing and Google recently,
as the find similar images function. The key component of image re-ranking is to compute
the visual similarities between images. Many image features have been developed in recent
years. However, for different query images, low-level visual features that are effective for
one image category may not work well for another. To address this, Cui et al. classified the
query images into eight predefined intention categories and gave different feature weighting
schemes to different types of query images. it was difficult for only eight weighting schemes
to cover the large diversity of all the web images. It was also likely for a query image to be
classified to a wrong category.Mapped visual features to a universal concept
dictionary.Lampert et al. used predefined attributes with semantic meanings to detect novel
object classes. Some approaches transferred knowledge between object classes by measuring
the similarities between novel object classes and known object classes (called reference
classes). All these concepts/attributes/reference-classes were universally applied to all the
images and their training data was manually selected. For each query keyword, a multi-class
classifier on low level visual features is trained from the training sets of its reference classes
and stored offline. If there are K types of visual features, one could combine them to train a
single classifier. It is also possible to train a separate classifier foreach type of features. Our
experiments show that the latter choice can increase the re-ranking accuracy but will also
increase storage and reduce the online matching efficiency because of the increased size of
semantic signatures.
ADVANTAGES
The query-specific visual semantic spaces can more accurately model the images to be reranked automatically.Visual features of images are then projected into their related visual
semantic spaces to get semantic signatures.Images are re-ranked by comparing their semantic
signatures obtained from the visual semantic space of the query keyword. An image may be
relevant to multiple query keywords. Therefore it could have several semantic signatures
obtained in different semantic spaces. According to the wordimage index file, each image in
the database is associated with a few relevant keywords. For each relevant keyword,
asemantic signature of the image is extracted by computing the visual similarities between
the image and the reference classes of the keyword using the classifiers trained in the
11
previous step. The reference classes form the basis of the semantic space of the keyword. If
an image has N relevantkeywords, then it hasN semantic signatures to be computed.
12
CHAPTER 4
SYSTEM DESIGN
4.1 ARCHITECTURE DIAGRAM
Keyword
Associated with images
DiscoveryOf
Query specific
Classifier
Reference
Reference
classes
Of
classes
signature over
reference classes
Reference classes
Offline
Online
Query Text
Text based image list
This diagram, Given a query key- word input by a user, according to a stored word-image
index le, a pool of images relevant to the query keyword are retrieved by the search engine.
By asking a user to select a query image, which reects the users search intention, from the
pool, the remaining images in the pool are re-ranked based on their visual similarities with
the query image. The visual features of images are pre-computed ofine and stored by the
search engine. The main online computational cost of image re-ranking is on comparing
visual features. Instead of constructing a universal concept dictionary, it learns different
visual semantic spaces for different query keywords individually and automatically. We
believe that the semantic space related to the images to be re-ranked can be signicantly
narrowed down by the query keyword provided by the user. For example, if the query
keyword is apple, the semantic concepts of mountains and Paris are unlikely to be
13
relevant and can be ignored. Instead, the semantic concepts of computers and fruit will
be used to learn the visual semantic space related to apple. The query-specic visual
semantic spaces can more accurately model the images to be re-ranked, since they have
removed other potentially unlimited number of non-relevant concepts, which serve only as
noise and deteriorate the performance ofre-ranking in terms of both accuracy and
computational cost. The visual features of images are then projected into their related visual
semantic spaces to get semantic signatures. At the online stage, images are re-ranked by
comparing their semantic signatures obtained from the visual semantic space of the query
keyword.
Several subsystems can be identified within the architecture of retrieval systems for images:
Pre-processing of the image to reduce the influence of different acquisition circumstances
(e.g., differences in illumination). Extraction of low-level features of visual data (e.g., Shape
and color).Image processing and pattern recognition is used to measure such features.
Extraction of high-level feature and image semantics (e.g., recognition of a shoe brand and
type based on the shoe profile).In some cases, semantics can be extracted automatically from
the images based on a combination of low-level features and rules. Description in textual
form of the image contents and acquisition characteristics (type of camera, image size, image
resolution, number of images in a sequence, information available about the individual
correspondence with the images, metadata describing the content etc.)Visualization, which
presents a view of the data for inspection, herby improving the effectiveness of the search.
Indexing and pre-selection, this filters out images that are not pertinent to a query and
extracts only those database items that are relevant to the query.
Retrieval: matching
procedures and similarity metrics. Relevancy feedback: a mechanism by which the user can
give feedback by indicating positive or negative relevance of retrieved items.
Functionality: Initially, user interface typically consists of a query formulation part and a
result presentation part. Specification of which images to retrieve from the database can be
done in many ways. One way is to browse through the database one by one. Another way is
to specify the image in terms of keywords, or in terms of image features that are extracted
from the image, such as a color histogram. Yet another way is to provide an image or sketch
from which features of the same type must be extracted as for the database images, in order to
match these features A combination of four feature extraction methods namely color
Histogram, Color moment texture, and edge histogram descriptor. There is a provision to add
new features in future for better retrieval efficiency. Any combination of these methods,
14
which is more appropriate for the application, can be used for retrieval. This is provided
through User Interface (UI) in the form of relevance feedback. Indexing is often used as
identifying features within an image; with indexing data structures we here mean structures to
speed up the retrieval of features within image collections.
4.2 DATAFLOW DIAGRAM
Start
Enter the
query
Query analysing by
image
Query grouping by
images
query relevance by
image
Flittering the
images
Graph
15
In this diagram, For each query keyword, a multi-class classier on low- level visual features
is trained from the training sets of its reference classes and stored ofine. If there are K types
of visual features, one could combine them to train a single classier. It is also possible to
train a separate classier for each type of features. Our experiments show that the latter
choice can increase the re-ranking accuracy but will also increase storage and reduce the
online matching efciency because of the increased size of semantic signatures. An image
may be relevant to multiple query keywords. Therefore it could have several semantic
signatures obtained in different semantic spaces. According to the word- image index le,
each image in the database is associated with a few relevant keywords. For each relevant
keyword, a semantic signature of the image is extracted by computing the visual similarities
between the image and the reference classes of the keyword using the classiers trained in the
previous step. The reference classes form the basis of the semantic space of the keyword. If
an image has N relevant keywords, then it has N semantic signatures to be computed and
stored ofine. At the online stage, a pool of images are retrieved by the search engine
according to the query keyword input by a user. Since all the images in the pool are relevant
to the query keyword, they all have pre-computed semantic signatures in the semantic space
of the query keyword. Once the user chooses a query image, all the images are re-ranked by
comparing similarities of the semantic signatures.
Indexing and Feature Extraction: Indexing deals with the insertion of feature vectors into the
database and is a fundamental task in every query-by-content database system. In many
applications, due to the diverse nature of queries, the feature vector may need to be
constructed from multiple, mostly unrelated, features. However, when diverse features are
present in a feature vector; it is very important how the feature space is organized. For
example, if two vectors are compared, which parts of the feature vector should be matched,
all or partial? Some feature may be of higher importance than the others. What is the
computational complexity of this matching? When all the features in a feature vector are
assumed to be equally important, the problem reduces to computing Euclidean distance in a
multidimensional space. In some other cases, different weights may be assigned to each of
the features in the feature vector, and the weights and the similarity metric are determined
through simulation based on certain optimization criteria.
16
In parallel to an efficient indexing system, a query system with the fast query response time is
very crucial in content- based image retrieval systems. the complexity of the computations
involved in query processing must be reduced as much as possible .since in content based
image indexing and retrieval systems images are indexed once(preprocessing stage) and
queried many times over and over, we may want to allocate complex operations to the
preprocessing stage.
17
CHAPTER 5
DEVELOPMENT REQUIREMENT
5.1 HARDWARE REQUIREMENT
Processor
Pentium III
Speed
1.1 Ghz
RAM
256 MB (min)
Hard Disk
20 GB
Floppy Drive
1.44 MB
Key Board
Mouse
Monitor
SVGA
Windows95/98/2000/XP
Application Server
Glassfish
Front End
Scripts
JavaScript.
Database
MySQL
Database Connectivity -
JDBC.
current version of Java is Java 1.7 which is also known as Java 7. From the Java
programming language the Java platform evolved. The Java platform allows software
developers to write program code in other languages than the Java programming language
and still runs on the Java virtual machine. The Java platform is usually associated with
the Java virtual machine and the Java core libraries.Java and Open Source: In 2006 Sun
started to make Java available under the GNU General Public License (GPL). Oracle
continues this project called OpenJDK. Java Virtual machine: The Java virtual machine
(JVM) is a software implementation of a computer that executes programs like a real
machine.
JSP technology is used to create web application just like Servlet technology. It can be
thought of as an extension to servlet because it provides more functionality than servlet such
as expression language, jstl etc. A JSP page consists of HTML tags and JSP tags. The jsp
pages are easier to maintain than servlet because we can separate designing and development.
It provides some additional features such as Expression Language, Custom Tag etc.
5.3.2 MySQL
MySQL is the most popular Open Source Relational SQL database management system.
MySQL is one of the best RDBMS being used for developing web-based software
applications. What is Database?A database is a separate application that stores a collection of
data. Each database has one or more distinct APIs for creating, accessing, managing,
searching and replicating the data it holds. Other kinds of data stores can be used, such as
files on the file system or large hash tables in memory but data fetching and writing would
not be so fast and easy with those types of systems. So nowadays, we use relational database
management systems (RDBMS) to store and manage huge volume of data. This is called
relational database because all the data is stored into different tables and relations are
established using primary keys or other keys known as foreign keys.
MySQL is a fast, easy-to-use RDBMS being used for many small and big businesses.
MySQL is developed, marketed, and supported by MySQL AB, which is a Swedish
company. MySQL is becoming so popular because of many good reasons: MySQL is
released under an open-source license. So you have nothing to pay to use it. MySQL is a very
powerful program in its own right. It handles a large subset of the functionality of the most
expensive and powerful database packages. MySQL uses a standard form of the well-known
SQL data language. MySQL works on many operating systems and with many languages
19
including PHP, PERL, C, C++, JAVA, etc. MySQL works very quickly and works well even
with large data sets. MySQL is very friendly to PHP, the most appreciated language for web
development. MySQL supports large databases, up to 50 million rows or more in a table. The
default file size limit for a table is 4GB, but you can increase this (if your operating system
can handle it) to a theoretical limit of 8 million terabytes (TB). MySQL is customizable. The
open-source GPL license allows programmers to modify the MySQL software to fit their own
specific environments.
20
The platform offers reusable services common to desktop applications, allowing developers
to focus on the logic specific to their application. Among the features of the platform are:User
interface management (e.g. menus and toolbars)User settings managementStorage
management (saving and loading any kind of data)Window managementWizard framework
(supports step-by-step dialogs)NetBeans Visual LibraryIntegrated development tools
NetBeans IDE is a free, open-source, cross-platform IDE with built-in-support for Java
Programming Language.
21
CHAPTER 6
SYSTEM IMPLEMENTATION
6.1 MODULES:
Login module.
Admin module
Text based image search.
Image re-ranking on one click.
Image retrieval
22
23
24
CHAPTER 7
25
REFERENCES
[1] E. Bart and S. Ullman. Single-example learning of novel classes using representation by
similarity. In Proc. BMVC, 2005.
[2] Y. Cao, C. Wang, Z. Li, L. Zhang, and L. Zhang.Spatial-bag-of- features.In Proc. CVPR,
2010.
[3] G. Cauwenberghs and T. Poggio. Incremental and decremental sup- port vector machine
learning. In Proc. NIPS, 2001.
[4] J. Cui, F. Wen, and X. Tang.Intentsearch: Interactive on-line image search re-ranking. In
Proc. ACM Multimedia.ACM, 2008.
[5] J. Cui, F. Wen, and X. Tang. Real time google and live image search re-ranking. In Proc.
ACM Multimedia, 2008.
[6] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection.In Proc.
CVPR, 2005.
[8] D. Lowe. Distinctive image features from scale-invariant keypoints. Intl Journal of
Computer Vision, 2004.
[9] B. Luo, X. Wang, and X. Tang. A world wide web based image search engine using text
and image content features. In Proceedings of the SPIE Electronic Imaging, 2003.
26
[11] N. Rasiwasia, P. J. Moreno, and N. Vasconcelos. Bridging the gap: Query by semantic
example. IEEE Trans. on Multimedia, 2007.
[12] M. Rohrbach, M. Stark, G. Szarvas, I. Gurevych, and B. Schiele. What helps wherevand
why? semantic relatedness for knowledge transfer. In Proc. CVPR, 2010.
[13] Y. Rui, T. S. Huang, M. Ortega, and S. Mehrotra. Relevance feed- back: a power tool for
interactive content-based image retrieval. IEEE Trans. on Circuits and Systems for Video
Technology, 1998.
[14] D. Tao, X. Tang, X. Li, and X. Wu. Asymmetric bagging and randomsubspaceforsupportvectormachines-basedrelevancefeedback
inimageretrieval.IEEETrans.onPatternAnalysisandMachineIn- telligence, 2006.
[15] Q. Yin, X. Tang, and J. Sun. An associate-predict model for face recognition. In Proc.
CVPR, 2011.
27
28