Dynamic Feature Selection and Coarse-To-Fine Search

for Content-Based Image Retrieval

J. You, Q. Li, K. H. Cheung

and Prabir Bhattacharya

Department of Computing, The Hong Kong Polytechnic University, KLN, Hong Kong

Institute for Information Systems Engineering, Concordia University

Montreal, Quebec, H3G 1T7, Canada

bstract. We present a new approach to content-based image retrieval by ad-

dressing three primary issues: image indexing, similarity measure, and search

methods. The proposed algorithms include: an image data warehousing structure

for dynamic image indexing; a statistically based feature selection procedure to

form ﬂexible similarity measures in terms of the dominant image features; and

a feature component code to facilitate query processing and guide the search for

the best matching. The experimental results demonstrate the feasibility and effec-

tiveness of the proposed method.

1 Introduction

The generation of large on-line collections of images has resulted in a need for new

methods which can 1) ﬁnd all images similar to a given image; 2) identify images

containing a speciﬁed object; 3) determine the derivatives of the original input image;

and 4) recognize images represented by speciﬁed colors, textures or shapes. In general,

image retrieval (IR) approaches fall into two categories: attribute-based methods and

content-based methods. Attribute-based methods represent image contents using text

and structured ﬁelds. Examples of such methods include Kodak Picture Exchange Sys-

tem (KPX) and PressLink library. However, the scope of applications of such methods

is limited because it is difﬁcult to describe image details such as irregular shapes and

jumbled textures in text.

Content-based image retrieval (CBIR) requires an integration of research in com-

puter vision, image understanding, artiﬁcial intelligence and databases to address the

problem of information retrieval from large collections of images and video frames. The

developments in this area are surveyed in [1], [2], [3], [4]. The CBIR approach is based

on the integration of feature-extraction and object-recognition during the management

of image databases to overcome the limitation of attribute-based retrieval. Traditionally

designing a CBIR technique involves three primary issues: image feature extraction

and representation, similarity measures, and image indexing and searching. Currently a

large proportion of research into CBIR is centered on issues such as which image fea-

tures are extracted, the level of abstraction manifested in the features, and the degree of

desired domain independence. Examples of CBIR systems include QBIC, Virage, Pho-

tobook, CANDID, Garlic, MetaSEEK, WebSeek and C-BIRD. Although the existing

You J., Li Q., H. Cheung K. and Bhattacharya P. (2005).

Dynamic Feature Selection and Coarse-To-Fine Search for Content-Based Image Retrieval.

In Proceedings of the 5th International Workshop on Pattern Recognition in Information Systems, pages 81-93

DOI: 10.5220/0002566800810093

 SciTePress

techniques resolve the design issues in various ways [2], they have the following lim-

itations: 1) the selection of image feature(s) for indexing and matching is pre-deﬁned

and lacks ﬂexibility; 2) the multiple features are not integrated for similarity measures;

3) the search for the best matching is computationally expensive and somewhat domain

dependent for object recognition by features. It is thus of signiﬁcant interest to develop

a general methodology for effective image feature extraction, indexing, searching and

integration.

Liang and Kuo [5] proposed a wavelet-based system, WaveGuide, which integrated

a set of image features including texture, color and shape in the wavelet domain for im-

age description and indexing. Although the system successfully handled the two compo-

nents of data representation and content description in a uniﬁed framework, it adopted a

ﬁxed image feature selection scheme and applied pre-deﬁned criteria for similarity mea-

sures. The recent MARS (Multimedia Analysis and Retrieval System) project [6] aims

to bridge the gap between the low-level features to high-level semantic concepts such

as complex objects and scenes by utilizing image understanding techniques. However,

the selection of features and indexing structure are ﬁxed in MARS.

To achieve ﬂexibility and multiple feature integration, we propose to construct a

decision tree for the retrieval task where a statistically based feature selection criterion

is used to guide the selection and integration of the most relevant features for similarity

measurement in a hierarchical structure.

Furthermore, we introduce a data warehousing structure and a feature component

code scheme to facilitate dynamic image indexing, multiple feature integration, ﬂexible

similarity measures and guided search.

This paper is organized in the following sections. Section 2 presents a dynamic

image indexing scheme which embraces multiple features in conjunction with the sta-

tistically based feature selection scheme and the feature component codes for large

collection of images based on the proposed image data warehousing structure. Section

3 describes a coarse-to-ﬁne search algorithm for image retrieval. The comprehensive

experimental results are reported in Section 4. Finally the conclusions are summarized

in Section 5.

2 Dynamic Image Indexing

2.1 Image Data Warehouse and Multi-dimensional Feature Cube

Unlike the traditional image databases which focus on only image data (content), the

proposed image data warehouse extends its capacity for data mapping and summariza-

tion. By mapping different image features to the data warehouse and generating sum-

mary tables based on the dimension data, the retrieval for similar feature data is guided

by directing a query to the right tables for fast searching. This search is much faster as

the size of these tables and the number of feature components is much smaller. Further,

as these tables are materialized views, information stored in them is pre-computed. The

reference to different summary tables will result in the selection of different indexing

schemes and similarity measurement. Thus, the use of the data warehouse will facili-

tate dynamic indexing to speed up the retrieval task. Fig. 1 illustrates the hierarchy of

multi-dimensional data cubes.

Summarized

Feature

Tables

Summarized

Feature

Tables

Summarized

Feature

Tables

SOURCE

(Shape)(Color) (Texture)

FEATURE REPRESENTATION

PHASE 1

PHASE 2

PHASE 3

Fig.1. A hierarchy of multi-dimensional data cubes

In our proposed image data warehouse system, the fact table deﬁnes the image con-

tent with three major image features: color, texture and shape. For each image feature,

the representation of the particular item is further described in the subsequent dimen-

sion table. Each of the ﬁelds in the current dimension table can be considered as facts

at a lower level and the details are summarized in the associated dimension table. Fig. 2

shows an example of such multi-level dimension tables for the representation of shape

feature, where Level 1 describes the major ﬁelds of an image content (fact table), Level

2 shows the further description of shape feature details in a corresponding dimension

table at a lower level (dimension table) and Level 3 lists the further details of the subse-

quent moment information for shape feature measurement at the next lower level. Note

that the amount of information stored in one of the lower tables is much less than in

the original fact table, which makes it easier to handle. For other image features such

as color, texture and salient feature points, the corresponding dimension tables can be

constructed in a similar manner.

With the proposed hierarchies of the image data warehousing structure, a multi-

dimensional feature cube can be used to visualize the image data warehouse and its

OLAP (On-line Analytical Processing) data, where the cells hold the quantifying im-

age feature components (often referred to as facts), while the qualifying feature indices

describe the axes of the cube and can be used for addressing individual feature com-

ponents or groups of multiple features. The proposed multi-dimensional structure of

the image feature cube offers ﬂexibility to access and manipulate the image data from

different perspectives. Such a structure allows quick data summarization at different

levels for dynamic image indexing and multiple feature integration. The statistical data

resulting from the OLAP operation is used to discover the hidden patterns or implicit

knowledge to speed up the task of CBIR. For example, a statistically based feature se-

lection criterion can be adopted to determine the importance of individual features in

indexing and similarity measurement. A comparison of our proposed approach with

the existing techniques is summarized in Table 1 and more details are reported in the

following sections.

2.2 Dynamic Feature Selection and Multiple Feature Integration

Most of the existing CBIR systems pre-deﬁne the features for the retrieval task in a

conventional database structure. In this paper, we propose to use a statistically based

CategoryImage ID

image #1

Color Texture Shape

texture dimension table

image #2

image #N

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Moment DescriptionB-spline Description Salient Points

B-spline coefficients

. . . . . .

moment feature vector

. . . . . .

feature point list

. . . . . .

Image ID

image #1

image #2

. . . . . .

LEVEL 1: The major fields of an image content (fact table)

LEVEL 2: The major fields of shape feature description (dimension table)

LEVEL 3: The data fields of invariant moments (dimension table)

(Order 0)

Invariant MomentInvariant Moment Invariant Moment

(Order 1) (Order 2)

moment measurement

moment measurement moment measurement

moment measurement

. . . . . . . . . . . . . . . . . .

image #1

image #2

. . . . . .

Image ID

feature component code

texture dimension tablecolor dimension table

color dimension table

shape dimension table

Fig.2. A hierarchy of multi-level dimension tables for shape representation

feature selection criterion to determine the most appropriate feature dimension table for

dynamic feature selection and indexing.

Recently feature selection criteria based on statistical measures have been exten-

sively studied. These include the Chi-square criterion, Asymmetrical Tau and Symmet-

rical Tau [9]. A comprehensive description and comparison of these criteria is given

in [9]. In this paper, we extend the use of the Symmetrical Tau criterion [9] to guide

the selection and combination of multiple image features for content-based image re-

trieval within the framework of image data warehousing structure. The key features

of our approach include: 1) apply Symmetrical Tau criterion to determine the impor-

tance of each individual image features; 2) use partitioning and aggregation technique

to generate the most appropriate feature dimension table for dynamic image indexing;

3) adopt a ﬂexible similarity measurement to integrate multiple features with different

weights. In addition, we apply a training procedure to determine the most appropriate

weights for multiple feature integration. Instead of using individual features to calcu-

late the corresponding Tau as initially deﬁned, we use the combined feature vector to

obtain the relevant Tau. Such a process is iteratively repeated by dynamically adjusting

the weights associated with each feature component. A higher weight is assigned to

the component for strong emphasis and a lower weight reﬂects less emphasis on the

non-relevant component. The set of weights with the maximum Tau is used to combine

multiple features for similarity measures. More speciﬁcally, we conduct the calculation

throughout the following major steps: 1) Identify all of the individual features to be used

for retrieval and obtain their feature vectors. For n features f

(i = 0, 1, ..., n − 1), there

will be n individual feature vectors V

(i = 0, 1, ..., n − 1). 2) Apply Gaussian normal-

Table 1. Comparison of different CBIR systems

Name of System Data Similarity Search Query

the System Structure Representation Measurement Method Process

QBIC [7] single image pre-deﬁned one-to-one user

database modeling measures comparison interactive

Garlic [8] multiple image&text fuzzy set object user

databases features technique oriented interactive

MARS [6] traditional image/video ﬁxed hierarchical user

database features measurement structure interactive

WaveGuide [5] traditional wavelet-based ﬁxed weighted user

database multiple features measurement ranking interactive

Mediahouse data multiple ﬂexible hierarchical user

(proposed approach) warehouse features measurement search interactive

ization to the above feature vectors. 3) Initialize a set of weights α

(i = 0, 1, ..., n − 1)

and obtain its corresponding combined feature vector V

n−1

i=0

. 4) Calculate

the corresponding Tau using the following formula:

T au =

j=1

i=1

P (ij)

P (+j)

i=1

j=1

P (ij)

P (i+)

− SUM

2 − SUM

(1)

where the contingency table has I rows and J columns; P (ij) is the probability that a

variable belongs both to row category and to column category j; P(i+) and P(+j) are

the marginal probabilities in row category i and column category j, respectively, and

SU M =

i=1

P (i+)

j=1

P (+j)

. 5) Adjust the set of weights, and obtain a new

combined feature vector V’

and calculate the corresponding Tau. 6) Repeat Step 5 for

all of the given adjustment weight sets. 7) Find the maximum value of Tau from the

sequences of Tau obtained in the previous stage. 8) Choose the combined feature with

the maximum Tau value.

3 Coarse-to-Fine Search

We propose a coarse-to-ﬁne search scheme in conjunction with feature component code

and selective matching criteria to speed up the process in a hierarchical fashion. Our ap-

proach is characterized as follows: 1) we use image feature component code to identify

the major feature components in an image; 2) we apply statistically based feature se-

lection criterion to rank the importance of each feature component and access to the

corresponding feature dimension table to group the potential candidates at coarse level;

3) we use different similarity measurements to conduct matching at the ﬁne level. Our

proposed system will support two types of queries: a) to pose a query by image sample,

and b) to use a simple sketch as a query.

In the case of query by image sample, the search follows multiple feature extraction

and image similarity measurement described in the previous sections. Based on the

nature of the query image, the user can add additional component weights during the

combination of image features for image similarity measurement.In the case of query

by a simple sketch provided a user, we will apply a B-spline based curve matching

scheme to identify the most suitable candidates from the image database. In the case

of a query by sample image, we will use the proposed image component code to guide

the search for the most appropriate candidates in terms of color, texture and shape from

data warehouse at a coarse level and apply image matching at ﬁne level for the ﬁnal

output. A fractional discrimination function is used to identify object boundaries for

coarse-to-ﬁne matching.

3.1 Feature Component Code

In this paper, we adopt a general wavelet based scheme for image feature extraction

and representation. Three features – color, texture and shape are considered and each is

associated with an individual feature vector. To organize the index structure with these

features to facilitate image query, we propose a three-bit feature component code to

characterize the status of each individual image, where the left bit C represents color

status (1 for a color image and 0 for a black and white image), the middle bit T rep-

resents texture status (1 for a texture image and 0 for a non-texture image), and the

right bit S represents shape status (1 for an image with clear objects and 0 for an image

without any objects).

Fig. 3 shows four image samples representing an outdoor scene, a texture, a sketch,

and a textured object respectively. The component code will facilitate the hierarchical

structure of indexing and searching. Each individual image is classiﬁed into different

image groups according to its component code. Within each group, images are further

ranked with respect to their individual feature vectors or measurements. The relevant

data is pre-processed and stored in the corresponding summary tables in the warehouse

structure. The search process will start with the image group which has the same compo-

nent code as the query image, which speeds up the processing by ﬁltering out irrelevant

images from the image collection.

(a) code 001 (b) code 010 (c) code 001 (d) code 011

Fig.3. Image feature component coding

3.2 A Multi-level Curve Matching Scheme

To avoid the blind searching for the best ﬁt between the template pattern and all of the

sample patterns stored in the image data warehouse, a guided search strategy is essential

to reduce computation burden. In a conventional data warehouse, various methods of

partitioning have been developed to improve the efﬁciency for query processing. Most

of the existing methods fall into two major categories – horizontal partitioning and

vertical partitioning. In general, the way in which a fact table will be split up depends on

the type of a query. In this paper, we adopted partitioning along a dimension to facilitate

a coarse-to-ﬁne curve matching scheme for similar shape retrieval. As illustrated in

Fig. 2, B-spline curve and invariant moments coefﬁcients are associates with the shape

feature in two dimensions and can be further partitioned into a lower level.

The initial search for the best similar shape matching starts with the dimension

of invariant moments. The similarity measurement is based on the comparison of the

Cosine distance of the proposed shape moment feature vectors for different samples

after Gaussian normalization. The candidates with small distance differences will be

considered for ﬁne matching based on 2D polygonal arc matching and B-spline curve

matching. The goal here is to match and recognize shape curves for the ﬁnal retrieval

output. The curve candidates which are selected from the 2D polygonal arc match will

be further compared based on their B-spline models. Traditionally the judgment is made

by the comparison of their control points, where the ordered corner points from bound-

ary tracing are used. The sample is allocated to the class with minimum difference

distance. However, the problem with this approach is that it cannot handle occluded

curves although it is straightforward to estimate the residual errors for similarity mea-

surements. Based on [10], we are able to use the Hausdorff distance algorithm to search

for portions, or partial hidden objects [11]. In the work reported here, we further extend

our previous research by using contour corner points as a basis for curve matching in

terms of the Hausdorff distance. It aims to reduce the computation required for a re-

liable match and be able to ﬁnd partially hidden curves. The Hausdorff distance is a

non-linear operator which determines the degree of the mismatch between a model and

an object by measuring the distance of the point of a model that is farthest from any

point of an object and vice versa. Therefore, it can be used for object recognition by

comparing two images which are superimposed on one another. For the curve match-

ing, set A and B contain the contour points of two curves respectively; and the best

matching occurs when its Hausdorff distance is minimal. To speed up the searching for

the best ﬁt between the two given curves, once again, we adopt a hierarchical scheme.

The Hausdorff distance at coarse level is determined by the B-spline control points of

each curve, and the more neighboring contour points are considered to determine the

Hausdorff distance at a ﬁne level.

4 Experimental Results

4.1 System Description

Our MediaHouse system applies data warehousing technique to handle important is-

sues of CBIR for a large collection of image samples from different sources. The 6,228

images used in our experiments include 3,834 palmprint samples of size 384 × 284 for

the case study of personal identiﬁcation via palmprint retrieval, 1,266 face images of

size 384 × 284 for the case study of face recognition based on similar shape retrieval,

and 1,128 natural images collected from various websites and public domains. The pre-

processing includes: 1) wavelet transforms of image samples; 2) multiple image feature

representation with respect to color, texture and shape; 3) generation of feature dimen-

sion tables. The Oracle data warehousing tool is used for data analysis and processing.

The operation of our system includes two basic phases: dynamic feature selection and

integrated similarity measures for phase I; and hierarchical search for the best match for

phase II. Our system supports query-by-example which compares the target image with

the possible candidates in the image data warehouse throughout the proposed guided

search procedure. Fig. 4 shows the diagram of the system structure.

Extraction

Transformation

Cleaning

DB1

DB2

DBn

Distributed

Multimedia

Databases

Data Marts

Inputs Outputs

Data

Multimedia

Multimedia Data Warehouse

OLAP Server

Decision Support

Data Mining

Modeling

Integration

Training

Fig.4. System structure of MediaHouse image data warehouse

4.2 Dynamic Feature Selection and Multi-level Similarity Measures

The dynamic selection of image features is demonstrated by multi-level palmprint fea-

ture extraction for personal identiﬁcation and veriﬁcation (see our previous work [12]).

The experiment is carried out in two stages. In stage one, the global palmprint features

are extracted at coarse level and candidate samples are selected for further process-

ing. In stage two, the regional palmprint features are detected and a hierarchical image

matching is performed for the ﬁnal retrieval. Fig. 5 illustrates the multi-level extraction

of palmprint features.

(a) palm (b) palm-boundary

Fig.5. Multi-level feature extraction

In our system, we consider multiple palmprint features and adopt different simi-

larity measures in a hierarchical manner to facilitate a coarse-to-ﬁne palmprint match-

ing scheme for personal identiﬁcation. Four palmprint features are extracted – Level-1

global geometry feature, Level-2 global texture energy, Level-3 local “interest” lines,

and Level-4 local texture feature vector. More speciﬁcally, the palm boundary segments

are used as Level-1 global geometry feature and they can be extracted by a boundary

tracking algorithm. The ‘tuned’ mask based texture energy measurement is used for

Level-2 global texture feature representation [12]. The dominant feature lines such as

principal lines, wrinkles, ridges in palmprint are extracted as Level-3 local ‘interest’

feature lines. A 2D Gabor phase coding is used to form Level-4 local texture feature

vector. We begin initial searching for the best similar palmprint matching group with

Level-1 global geometry feature. Our similarity measurement method is based on the

comparison of the boundary segment with respect to its length and tangent for different

samples. The candidates with small distance differences will be considered for further

coarse-level selection by global texture measurement. The selected candidates will be

subjected to ﬁne matching based on 2D polygonal arc matching and comparison of local

texture feature vector.

The proposed ﬁne matching algorithm starts with 2D polygonal arc matching by

using a subgroup of the unit quaternions. For palmprint samples which consist of the

best matched arcs will go through the ﬁnal ﬁne matching in terms of their local texture

feature vectors. The ﬁnal matching is based on the comparison of local texture feature

vectors via 2D Gabor coding. The best match is the candidate with the least normalized

Hamming distance.

The palmprint image samples used for the testing are size of 384× 284 with the res-

olution of 75 dpi and 256 gray scales. In our palmprint image database, 3834 palmprint

images from 193 individuals are stored. These palmprint samples are collected from

both female and male adults with the age range from 18 to 50. A series of experiments

have been carried out to verify the high performance of the proposed algorithms.

The veriﬁcation accuracies at different levels are shown in Fig 6. Fig 6 A-1,A-2,

A-3 and A-4 present the probability distributions of genuine and imposter at different

feature levels. The corresponding receiver operating curves (ROC), being a plot of

genuine acceptance rate against false acceptance rate for all possible operating points

are demonstrated in Fig 6 B-1, B-2, B-3 and B-4. Based on ROCs, we conclude that

Level-4 local texture feature performs better than our hierarchical palmprint approach

when the false acceptance rate is large, such as 5%. However, a biometric system should

always operate in the condition of low false acceptance rate. Therefore, our hierarchical

palmprint approach is better than ﬁne-texture with low false acceptance rates. The false

rejection and correct rejection rates of the ﬁrst three levels are given in Table 2. The

improvement of system efﬁciency by the use of hierarchical scheme is illustrated in

Fig. 7 of system performance.

4.3 Multi-level Similar Shape Retrieval

The problem of similar-shape retrieval concerns retrieving or selecting all shapes or

images that are visually similar to the query shape or the query image’s shape. The

proposed coarse-to-ﬁne curve matching approach is demonstrated in the test of face

A. Genuine and imposter distributions for different features

(a) Level-1 (b) Level-2 (c) Level-3 (d) Level-4

B. Receiver operator curve (ROC) distributions for different features

(e) Level-1 (f) Level-2 (g) Level-3 (h) Level-4

Fig.6. Palmprint veriﬁcation test using multi-level features

Table 2. The system performance of accuracy

False Reject Rate Correct Reject Rate

1.69% 92.47%

1.10% 86.62%

0.73% 79.27%

recognition for personal identiﬁcation. At the coarse level, a fractional discrimination

function is used to identify the region of the interest of an individual’s face. At ﬁne curve

matching level, the active contour tracing algorithm is applied to detect the boundaries

of interest face regions for the ﬁnal matching. Fig. 8 illustrates the tracing of face curves

for face recognition. Fig. 8(a) is an original image, Fig. 8(b) shows the boundaries of

interest regions on the face and Fig. 8(c) presents the curve segments for hierarchical

face recognition by curve matching.

To verify the effectiveness of our approach, a series of tests are carried out in a data-

base of 1,266 face collections from different individuals under various conditions such

as uneven lighting, moderate tilting and partial sheltering. Fig. 9 shows the detected

regions of different individuals detected by the proposed algorithm and Table 3 lists the

correctness rate of the coarse-level detection.

To show the robustness of the proposed algorithm for face detection invariant of per-

spective view, partial distortion and occlusion, the ﬁne-level curve matching is applied

Fig.7. The system performance

(a) original image (b) boundary detection (c) face curves

Fig.8. Face curve extraction

to face images with different orientations and expressions. Fig. 10 illustrates face sam-

ples of the same person at various perspective views and Table 4 summarizes the testing

result for 100 testing cases. Fig. 11 shows face samples of the same person at different

conditions such as face expression, partial occlusion and distortion. The testing result

for 100 cases is listed in Table 5.

5 Conclusion

Image feature extraction, indexing and search are essential issues in content-based im-

age retrieval. In contrast to the existing techniques which mostly apply a ﬁxed mecha-

nism for feature extraction and indexing, we have proposed a statistically based feature

Fig.9. The detection of face region of interest

Table 3. Performance of face detection at coarse-level

Face Condition Correct Detection Rate

unevenness of lighting 98%

multiple faces 95%

moderate tilt of faces 97%

partial sheltering 85%

Fig.10. The face samples at different orientations

selection scheme for dynamic multiple feature extraction and integration within the im-

age data warehousing structure. This method is characterized as a ﬂexible but simple

approach to extract multiple features for similarity measures. In addition, the introduc-

tion of fractional discrimination function is powerful for multi-level feature extraction

and the proposal of image feature component code facilitates guided search for the best

matching. The experimental results provide the basis for the further development of

an effective and efﬁcient content-based image retrieval system. For future research, it

would be of considerable interest to investigate the following issues. Firstly, we would

like to develop a comprehensive interactive query interface with the full support of user

feedback. Secondly, we would like to include more image samples of diversity in our

image data warehouse and apply data mining techniques for image understanding and

feature clustering. Thirdly, we would like to develop a general scheme for performance

evaluation in terms of accuracy and efﬁciency in addition to the current measurement

of precision and recall.

Acknowledgment

The support from the Hong Kong Polytechnic University Research Grants is greatly

appreciated.

Table 4. Performance of face recognition at different orientations

Viewing Perspective Correct Classiﬁcation Rate

−20

(vertical) 84%

−10

(vertical) 86%

+10

(vertical) 86%

+20

(vertical) 83%

−20

(horizontal) 85%

−10

(horizontal) 87%

+10

(horizontal) 87%

+20

(horizontal) 84%

Fig.11. The face samples at different conditions

Table 5. Performance of face classiﬁcation with different conditions

Face Condition Correct Classiﬁcation Rate

partial occlusion 77%

various expressions 81%

wearing glasses 82%

References

1. Antani, S., Kasturi, R., Jain, R.: A survey on the use of pattern recognition methods for

abstraction, indexing and retrieval of images and video. Patt. Recog. 35 (2002) 945–965

2. Smeulders, A., Santini, M., Gupta, A., Jain, R.: Content-based image retrieval at the end of

the early years. IEEE Trans. Patt. Anal. Machine Intell. PAMI-22 (2000) 1349–1380

3. Rui, R., Huang, T., Chang, S.: Image retrieval: current techniques, promising directions, and

open issues. J. Visual Communication and Image Representation 10 (1999) 39–72

4. Shapiro, L., Stockman, G.: Computer Vision. 2nd edn. Prentice-Hall, Englewood Cliffs, NJ,

N.J. (2001)

5. Liang, K., Kuo, C.: Waveguide: A joint wavelet-based image representation and description

system. IEEE Trans. on Image Processing 10 (1999) 39–72

6. Huang, T., Naphade, M.: Mars (multimedia analysis and retrieval system): A test bed for

video indexing, browsing, searching, ﬁltering, and summarization. In: Proc. 2000 IEEE Int.

Workshop on Multimedia Data Storage, Retrieval, Integration and Applications, Hong Kong

(2000) 1–7

7. Flickner, M., Sawhney, H., Niblack, W., Ashley, J.: Query by image and video content: The

qbic system. IEEE Computer 28 (1995) 23–32

8. Cody, W., Haas, L., Niblack, W., Arya, M., Carey, M., Fagin, R., Flickner, M., Lee, D.,

Petkovic, D., Schwarz, P., Thomas, J., Williams, J., Wimmers, E.: Querying multimedia data

from multiple repositories by content: the garlic project. In: Proc. of IFIP 3rd Working Conf.

on Visual Database Syst. (VDB-3). (1995)

9. Zhou, X., Dillon, T.: A statistical-heuristic feature selection criterion for decision tree induc-

tion. IEEE Trans. Patt. Anal. Machine Intell. PAMI-13 (1991) 834–841

10. Huttenlocher, D., Klanderman, G., Rucklidge, W.: Comparing images using the hausdorff

distance. IEEE Trans. Patt. Anal. Machine Intell. PAMI-15 (1993) 850–863

11. You, J., Bhattacharya, P.: A wavelet-based coarse-to-ﬁne image matching scheme in a paral-

lel virtual machine environment. IEEE Trans. on Image Processing 9 (2000) 1547–1559

12. You, J., Li, W., Zhang, D.: Hierarchical palmprint identiﬁcation via multiple feature extrac-

tion. Patt. Recog. 35 (2002) 847–859