Automatic Building of a Visual Interface for Content-based Multiresolution Retrieval of Paleontology Images J´erˆome Landr´e¦ , Fr´ed´eric Truchetet¦ , Sophie Montuire? , Bruno David? ¦ Institut
Universitaire de Technologie, Le2i, FRE 2309 CNRS, 12, rue de la Fonderie, 71200 Le Creusot, France
? Universit´ e
de Bourgogne, laboratoire de Pal´eontologie, UMR 5561 CNRS, 6, boulevard Gabriel, Dijon, France e-mail:
[email protected]
ABSTRACT This article presents a research work in the field of content-based image retrieval in large databases applied to the Paleontology image database of the Universit´e de Bourgogne, Dijon, France called “TRANS’TYFIPAL”. Our indexing method is based on multiresolution decomposition of database images using wavelets. For each family of paleontology images we try to find a model image representing it. The K-means automatic classification algorithm divides the space of parameters into several clusters. A model image for each cluster is computed from the wavelet transform of each image of the cluster. Then a search tree is built to offer users a graphic interface for retrieving images. So that users have to navigate through this tree of model images to find an image similar to that of their request. Our contribution in the field is the building of the model and of the search tree to make user access easier and faster. This paper ends with a conclusion on experimental results and a description of future work to be done to enhance our indexing and retrieval method. Keywords: Content-based Image Retrieval, Clustering, Multiresolution Analysis, Wavelets, K-means Automatic Classification, Paleontology Images Databases, TRANS’TYFIPAL
1. INTRODUCTION Due to the exponential growth of the Internet, a lot of multimedia databases became available online proposing texts, images, videos and sounds to web surfers. The problem of retrieving information from these large bases quickly appeared. How to find a specific multimedia resource from a large database with a good response time ? In our work, we try to answer this question for still images only from a single image database (it is already a large problem). The first methods used were based on keywords associated with images. The association was too subjective to give good results. The work of indexing each image with keywords was easy with small databases but is really unthinkable with large ones. Many methods to try to solve this problem exist. They are generally based on region segmentation, they use local, global or both local and global features computed from the images. Content-based image retrieval (CBIR) consists in retrieving an image without any “a priori” knowlegde about it. We only use the image itself to match a user request. This request can take several forms: a similar image, a hand-painted approximation (query by example), even a written description of the image. In the latter case, we are faced with the same problem as keywords indexing because the description a user gives is too subjective. For our research work, an image database from the paleontology laboratory of the Universit´e de Bourgogne, Dijon, France was used. This database demo is available online on the web at the Universit´e de Bourgogne website “http://www.u-bourgogne.fr/BIOGEOSCIENCE/ttf2.html” ∗ . When up-to-date, this database will contain about 60,000 paleontology images of a lot of animal and vegetal species from the beginning of life on Earth to the present day. ∗
Your computer needs to be Domain Name Service (DNS) declared to access this site.
A review of previous related papers is given in section 2. General methods are presented in section 2.1 while section 2.2 deals with wavelet-based techniques and section 2.3 insists on clustering techniques. Our decomposition method using wavelets is explained in section 3. In the section 3.1, the algorithms are detailed. Section 3.2 shows the transformations applied to each image of the database. Section 3.3 gives the indexing technique for transformed images. The method for building a model image is introduced in section 3.4. The retrieval technique employed is described in section 3.5. Section 4 explains the application of our method an a specific database and section 5 gives experimental results. Finally, section 6 gives a conclusion and describes future work to do about our method.
2. STATE OF THE ART 2.1. General Methods Several methods for indexing and retrieving still images from large databases have been published. Different methods give different results. There is no unified technique to test the different existing methods because each method is tested on a different database. The unique way of testing should be to create a large database with a lot of different images and to propose it on the web. Then each method could be tried on this base in order to compare image retrieval systems. The project was proposed in the year 2000 and is still in construction. 1 It is very difficult to give a complete state of the art for image retrieval systems because of the number of papers published each year about this research field. Rui2 et al. gave a good survey of this field. In CBIR systems, the general method consists in computing a n-dimensional feature vector V i representing each image i and computing a feature vector representing the request Vr . And then evaluating a distance between each Vi and Vr in order to sort the images and to give a result to users. Many distance functions were published, giving many different results. A list of some image retrieval systems was given by Peˇcenovi´c.3 The most popular CBIR system is QBIC4,5 from IBM. The image indexing and retrieval technique is based on texture indexing of image regions. Another system Photobook6 from MIT Media lab is very efficient for homogeneous image databases. It uses color, texture and shape features to work. In the Virage system7 color location and texture can be used with various weights in order to find a set of images similar to a request. We could also cite the Chabot system, 8 VisualSeek,9 RetrievalWare10 and Netra.11 All these systems are easy to use but need several parameters chosen by the user when accessing the database. Another approach is to work on the user interface to access and search the base. Chen, Bouman and Dalton 12 have already worked on building an easy-to-use interface for database browsing. They use the K-means algorithm to build a quadtree to help users to search the database easily, but they do not use wavelets for their work and our approach is slightly different. This paper describes our method based on wavelet-based multiresolution decomposition of database images.
2.2. Wavelet-Based Methods Our technique uses a multiresolution decomposition of images using wavelets. Wavelets were first introduced by Grossman13 and Morlet as a mathematical tool for analyzing seismic signals. Then the theory was developed by many contributors.14–16 Today, wavelets are a very powerful mathematical tool used for many signal processing tasks17 . Mallat18 showed that wavelets were a very good tool to detect singularities into a signal. Our method is based on this remark: we use wavelets to characterize the signal by representing it with some parameters computed from its multiresolution analysis. Several wavelet-based methods for CBIR have been published. In their paper, Jacobs, 19 Finkelstein and Salesin use multiresolution analysis to create an index generated from the indices and signs of the largest-magnitude wavelet coefficients. So they compare the index of the request and the database images using a distance function. Another method based on moments and wavelets was proposed by Mandal,20 Aboulnasr and Panchanathan. Histograms of the wavelet subbands are used for indexing. Idris21 and Panchanathan give a method based on wavelets vector quantization in which they compare feature vectors built upon wavelet coefficients. The idea of progressive retrieval is given by Liang22 and Kuo. Wavelet coefficients are computed at several levels and then they go through each level during retrieval. Stark23 describes a multiresolution technique using neural networks for wavelet-texture analysis classification. In his report, Peˇcenovi´c24 details a wavelet-packet algorithm to approximate Karhunen-Lo`eve Transform on images to index and retrieve them. Do25 proposes a wavelet maxima moment method where wavelet
decomposition is followed by indexing locally maxima wavelet moments. Chen, 26 Li and Chien work with color image segmentation at multiple scales allowing progressive retrieval. Our previous paper 27 gave a short introduction about our method.
2.3. Clustering Methods Many clustering methods28–31 have been proposed in the past. The idea is to divide a parameter space into several clusters of images and to represent each of them with a single model image. Different techniques are used in order to classify images into clusters. In Sheikholeslami32,33 et al. approach, a request icon image is compared to a cluster icon at a coarse resolution. Then the retrieving process goes on with only the matching cluster using fine resolution. This allows a rapid selection of a cluster before going on with the research process. Those methods were able to use pre-selection of clusters before continuing the retrieval process but they still used image comparison. In our approach, There is no comparison between a request image and the database images. The features are divided automatically into clusters. A model image is computed to represent each cluster and it is proposed visually to the user who chooses the closest image to the searched image. The user knows what kind of image he or she needs and select visually the closest image to his or her idea. Then the process goes on in the chosen cluster which is sub-clustered again and again to finally propose to the user only a few set of images as a result. Non-wavelet methods are not so flexible than wavelet-based methods because in the latter, multiresolution representation of data can lead to multiresolution classification of data. This gives a coarse-to-fine approach for the classification which is very important when the size of the database increases. Wavelet-based techniques increases the flexibility and the speed of the entire retrieval process. However wavelet transform is a time consuming task that needs to be performed off-line. As far as we know, none of the previous wavelet decomposition techniques use automatic classification of images in order to build a search tree. Our solution helps users to navigate through the database quickly and easily by choosing the image closest to their request at each level of the tree with the ability to go back and change their choice.
3. MULTIRESOLUTION DECOMPOSITION METHOD 3.1. General overview The multiresolution decomposition technique using wavelets was introduced by Mallat 15 and gives — from an original image — an approximation image and one or more detail image(s) at multiple scales. Each scale contains details of the original image. As Mallat18 pointed out, wavelets are good to detect singularities in signals. Edge detection and texture discrimination can be computed directly from wavelet coefficients in order to build an index. An important factor to take in consideration when working with database is the speed of the process because the time of computing increases critically with the size of the database. Multiresolution decomposition of images consists in filtering images with a row filter followed by a column filter. So the speed depends on the size of the filters used. The size of the filters in our case does not need to be very large because we only want to detect singularities in images. For instance, the Haar filter — whose size is only two taps for rows H haar = (1/2, 1/2) and two taps for columns Ghaar = (−1/2, 1/2) — could be sufficient for analyzing images. The choice of the wavelets will be discussed further in this paper. Multiresolution decomposition offers four interesting points for our work: • Multiresolution decomposition: each image is decomposed into an approximation image and one or more detail image(s) at multiple scales giving us the opportunity to find good parameters representing each family in our database at multiple scales (used for indexing). Each different scale gives different details leading us to use progressive retrieval methods. The low resolution approximation image can be used as an icon to represent the original image, • Several wavelet-based compression techniques have been studied in the recent years allowing us to compress our image to reduce our database size,
• Image watermarking is possible using wavelet decomposition and low-level details information. We can make our images watermarked for copyright reasons, • Reconstruction is possible allowing us to store transformed images directly in the database instead of original images. This can be a solution to save computing time while accessing our database. In this article, we will focus on the first abovementioned point, leaving compression, watermarking and reconstruction for future work. An image is represented at multiple resolutions, each resolution giving fewer details than the previous one. Figure 1 gives a pyramid where the original image A0 is the root. A0 is transformed into an approximation image A1 and one or more detail image(s) D1. The process is repeated and A1 is transformed into A2 and D2. At the end of the process, A3 is the approximation image and D3 is (or are) the detail image(s) of our original image. A0
D1
A1
A2
D2
A3 D3
Figure 1. An overview of multiresolution decomposition The proposed CBIR system is given on figure 2. The system is divided into two parts: the CBIR system itself which computes everything and builds the tree and the user interface which is the search tree. The database can only be accessed by user via the search tree. The principles of the method are simple. All the images of the database are transformed to a prefixed level of decomposition L using wavelets. Several features are extracted from these transformed images at multiple resolutions. Once put together, the features are used as the index of size n for the images. Then a classification is performed automatically using the K-means algorithm. The number of classes K after the classification is a parameter of the CBIR system. A representative image — called the model image — is built for each family constructed by the classification process. These model images are put together in a balanced K ary search tree. The number of nodes at each level of the tree is K. The tree has L levels in depth. The user can access the database only with the search tree by choosing the closest image to the request. The user uses visual similarity to navigate through the tree.
3.2. Transforming images Let us define several notations and describe the different algorithms. All the images of the base have to be square matrices for computing convenience. Only grayscale images are used to simplify and accelerate the process. Let D
Wavelet transform
Feature extraction
Images
Automatic Classification
Database manager
Model and tree building
Research tree
Result
User
Figure 2. Overview of the proposed CBIR system
be the database, i be an image, ig be the gray-level image of i, isq the square image representing i, iwt the wavelet transform of i and let w ˜ be the wavelets used for the analysis. The first step of our method is to compute and store the wavelet transform of each image.
f or each i ∈ D { load i ig ← grayscale − image(i) isq ← resize(ig , 256, 256) iwt ← wavelet − decomposition(isq , w) ˜ save iwt }
3.3. Indexing After this part of the process, the transformed images are stored for further reference. The next part of the work is to build an index. The purpose of indexing is to speed up the process of retrieving images. Instead of working directly with images (a long and fastidious task), the idea is to use a reduced set of parameters characterizing the images. For each wavelet transformed image n parameters are computed and normalized to belong to [0,1]. So a set of parameters (the index) pˆn ∈ [0, 1]n is obtained. The choice of the parameters inside the index is critical in CBIR systems since they characterize images. The computed parameters are based on wavelet coefficients for each detail image: maximum value, minimum value, mean value, normalized energy, histogram maximum value, histogram minimum value, histogram mean value. . . can be computed easily at each scale.
f or each iwt { load iwt pn ← set − parameters(iwt , n) pˆn ← normalize(pn ) save pˆn }
3.4. Building a model With the index made of normalized parameters, the tree can be built automatically using the K-means algorithm. For each level of the tree, all the images are classified in the hyperspace of parameters. The K-means algorithm automatically generates a set of classes from a parameter space by using the following algorithm. The goal is to create automatically K classes (clusters) from a space of vectors made of n normalized parameters.
randomly generate K vectors {Vr ∈ [0, 1]n ; r ∈ [1, K]} repeat { f or r = 1 to K Cr ← {v : d(v, Vr ) < d(v, Vj ) f or j 6= r} f or r = 1 to K Vr ← centroid(Cr ) } until centroids are stable
The algorithm classifies each image into K families at each level of the tree to obtain a K ary tree. An image is qP n 2 then labeled with the number of its class. The Euclidean distance d(u, v) = j=1 (uj − vj ) was chosen because of its isotropic properties.
load pˆn f or j = 1 to L { imagesj ← K − means(ˆ pn (j), K) repj ← mean(10% imagesj ) save repj }
In each class, the model image of this class is the mean image (repj ) of the 10% of images that are closest to the centroid of the class if the class contains more than ten images and the mean of the images of the class otherwise. So K model images representing K classes are used at the root of the tree. At the next level, for each class, another classification is performed. K other images are obtained which are the model for the next level. The process is reiterated to build the tree.
3.5. Retrieving images When searching our database, a user goes through the built tree, clicking images similar to what he or she wants. The K root models of our tree are proposed to the user who chooses one of them. Again, K model images of the previous class are shown. The user can choose and navigate through these images to find what he or she is looking for. The leaves of the tree are the original images of the database. Here is an abstract of the full process: For each image in the database, a wavelet transform is computed. This leads to the extraction of several parameters at multiple scales. The index is created from these parameters by normalization. Then the images are classified automatically into clusters using the K-means algorithm. A search tree — in which the model images are the mean of the 10 % closest images to the centroid of the cluster — is built. The user navigates the database by clicking the images of the tree, looking for an image similar to what he or she is looking for.
4. TESTING THE METHOD A paleontology image database called “TRANS’TYFIPAL” is used to test our method. Several images extracted from this base are shown in figure 3. This database covers a lot of paleontology species giving a wide range of images.
Figure 3. Several examples of TRANS’TYFIPAL images. In this article, only gray-level images are used for three reasons: • “TRANS’TYFIPAL” images often come from the earth or far down the ocean, so their colors have been altered by millions of years under the ground or under water,
• Due to a bad conservation state, some of the specimens were artificially colored before photography, • A red ammonite is an ammonite. The color of the specimen does not matter for our study. So the method was tested using the previous remarks. A first set of 123 gray-level images was used to test the method. A dyadic separable wavelet transform was used to analyze our database. In this decomposition, each transformed image is a quarter of the size of the upper level image. So we have to work with square decomposable image sizes, for instance 512x512, 256x256, 128x128. . . This kind of transform offers an approximation image and three detail images: horizontal, vertical and diagonal. An example of this kind of decomposition is shown in figure 4.
50
100
150
200
250 50
100
150
200
250
Figure 4. An example of 3-level multiresolution analysis on a TRANS’TYFIPAL image. Horizontal, vertical and diagonal details are visible at each different scale. The algorithm is simple as it works with square constant-sized images. For each image, a three-level wavelet transform is computed to go from 256x256 original image to 64x64 low-level image. Haar wavelets and spline wavelets are the wavelet families chosen. The former because they are fast and easy to compute, the latter because their localization is better. Then for each level and for each detail image, five normalized parameters are computed: minimum, maximum, mean, energy, normalized energy of wavelet coefficients at each scale. A vector made of fourtyfive parameters is obtained. This vector is the index for our database. So a binary tree K=2 at L=3 levels in resolution is computed.
5. EXPERIMENTAL RESULTS As a result, figure 5 shows an example of a three-level classification using the K-means algorithm. On figure 6, the corresponding model images are shown for each node of the tree. After that, this binary tree is proposed at the user in a browsing environment (like a web client) to search the database with the opportunity to go back if the choice is not good. As the computing time is very important in content-based retrieval systems, figure 7 gives the computing time of the algorithm for a second set of 246 images for an increasing number of parameters extracted from the wavelet coefficients. The computer used was an Intel Pentium II 266 MHz processor with 128 Mb SDRAM running Microsoft Windows 98 and Matlab 5. These computing times includes only the classification, building of the model and building of the tree times. The first phase of resizing, gray-level transforming and the wavelet transform was on average about 5.6 s per image. The second phase i.e. the computing of index parameters was about 1.1 s per image. The time necessary for the computation of the tree and the model images (in seconds) is given for a 246 images database on figure 7. In order to illustrate our study, only two parameters were used: the minimum value and the
83−40 1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
35−48
12−28
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
13−22
0
1
0
0.1
31−17
0.2
0.3
0.4
0.5
0.6
0.7
1
1
0.9
0.9
0.8
0.8
0.8
0.8
0.7
0.7
0.7
0.7
0.6
0.6
0.6
0.6
0.5
0.5
0.5
0.5
0.4
0.4
0.4
0.4
0.3
0.3
0.3
0.3
0.2
0.2
0.2
0.2
0.1
0.1
0.1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
1
0
0.1
0.2
0.3
0.4
0.5
1
13−15
1
0.9
0
0.9
6−6
1
0.9
0
0.8
0.6
0.7
0.8
0.9
0
1
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 5. Images classification using the K-means algorithm. The black circles are the centroids of each class. The number of images in each class is shown at the top of each graph. cg3−9−83
cd3−4−40
50
50
100
100
150
150
200
200
250
250 50
100
150
200
250
50
cg3−9−83g2−4−35
cg3−9−83d2−5−48
100
150
200
250
cd3−4−40g2−2−12
cd3−4−40d2−3−28
50
50
50
50
100
100
100
100
150
150
150
150
200
200
200
250
250 50
100
150
cg3−9−83g2−4−35g1−2−13
200
50
cg3−9−83g2−4−35d1−3−22
200
250
250
100
150
200
250
250
cg3−9−83d2−5−48g1−4−31
50
cg3−9−83d2−5−48d1−2−17
100
150
200
cd3−4−40g2−2−12g1−6−6
250
50
cd3−4−40g2−2−12d1−6−6
100
150
200
cd3−4−40d2−3−28g1−2−13
cd3−4−40d2−3−28d1−2−15
50
50
50
50
50
50
50
50
100
100
100
100
100
100
100
100
150
150
150
150
150
150
150
150
200
200
200
200
200
200
200
250
250 50
100
150
200
250
250 50
100
150
200
250
250 50
100
150
200
250
250 50
100
150
200
250
250 50
100
150
200
250
200
250 50
100
150
250
200
250
250 50
100
150
200
250
50
100
150
200
250
Figure 6. Model images of each class. This tree is the binary tree proposed to users in order to search the database.
Building time for the tree 1 level 2 levels 3 levels 4 levels 5 levels
60
Time (s)
50 40 30 20 10 0 0
20
40
60
80
100
Number of parameters Figure 7. Building time for several levels of the tree and several parameters.
normalized energy of the first-level coefficients. The times indicated were computed once, their values depend of the randomly generated kernel of the K-means algorithm. The number of parameters does not change critically the computing time of the algorithm. It means a lot of parameters can be used to characterize our images without having a costing time implementation. This was not an “a priori” result before the test of the method. Figure 8 shows the computing time for several database sizes with n = 100 parameters as the size of the index.
Building time for the tree 1 level 2 levels 3 levels 4 levels 5 levels
200
Time (s)
150
100
50
0 0
200
400
600
800
1000
1200
Database size Figure 8. Building time for several levels of the tree and several database sizes. As already mentioned, computing time increases critically with the size of the database. But better performances could be obtained with a more recent computer with a more powerful processor and more memory. We can imagine
to run the classification at one level in depth rather than multiple levels. So that the building of the tree could be in “real time” at each choice of the user. This paper does not present the classical precision and recall graphs because the K-means algorithm gives different clusters for different randomly chosen kernels. So testing the method twice will not lead to the same clusters, precision and recall graphs are not accurate for our study. The purpose of the method is not to get real families (clusters) of paleontology images, but clusters with visually similar images. An example of a classification is given in figure 9. The images are represented with respect to their real space position (projected onto a 2d space) to the center of the class. As already mentioned, the purpose of our method is to group visually similar images into clusters. On this figure, several images are grouped together into a cluster. The representative image is built with the three closest images to the center of the cluster. This example illustrates the kind of obtained groups at low resolution where only the coarse features (general shape) of the images are used to make the clustering. The group taxonomy takes in account texture, fine details, shape details. . . when increasing the resolution.
6. CONCLUSION In this paper, a method for visual interface building has been introduced. As far as we know, the idea of proposing a search tree based upon wavelet decomposition of images is new. The building of a model image allowing the user to navigate the database by clicking simplifies the approach of CBIR systems. As shown in the example, the method gives good results to classify images at each level of the tree. The computing time of the method is low, even if the database size and the index size increase. A lot of studies are possible to enhance the results. An interesting study would be to test the method on a general image database to see the results. Several different wavelet families could be used to detect different singularities. Different color spaces typically RGB and HSV could give better results than grayscale images. During the creation of the model, a weighted parameter depending on the distance of the representative images to the centroid of the cluster could be used i.e. the more the distance to the centroid the less the image is important in the model. The tree could have the possibility to be unbalanced. For instance, if the classification gives one very large family and another very small, the process could go on with subclassing the big family and stop for the small one. An inter-class distance function could be used to allow some images to belong to more than one family to add fexibility to the system. This method gives the opportunity to extract parameters at each scale of the decomposition. These parameters are specific to each scale, so in the index, different sizes of parameters could be use at each resolution allowing to use a progressive retrieval scheme. The decomposition goes from high levels to low levels. When searching the database, the low-level indexes (lowest index size) are used first. Then, the more we want to refine, the higher the level. With this way of searching, one can force a search time for our system. If the user wants quick response then only low-level indexes are used, if not, (time consuming) high level indexes are used. A lot of work has to be done to prepare future wavelet-based CBIR systems.
ACKNOWLEDGMENTS All the images in “TRANS’TYFIPAL” and in this document are the property of the laboratoire de Pal´eontologie de l’universit´e de Bourgogne, Dijon, France. Thanks to Fran¸coise Gasquez for her help about the “TRANS’TYFIPAL” database.
REFERENCES 1. http://www.benchathlon.net. 2. Y. Rui, T. S. Huang, and S.-F. Chang, “Image retrieval: Current techniques, promising directions and open issues,” Journal of Visual Communication and Image Representation 10, pp. 1–23, 1999.
Representative image ¡ ¡ ª
¡
Figure 9. An example of a classification of a cluster made upon 29 images. This example is taken from another classification of the same base and shows the visual similarity of the images.
3. Z. Peˇcenovi´c, M. Do, S. Ayer, and M. Vetterli, “New methods for image retrieval,” in Proceedings of the International Congress on Imaging Science, vol. 2, pp. 242–246, (University of Antwerp, Belgium), September 1998. 4. M. Flickner, H. Sahwney, W. Niblack, J. Ashley, Q. Huang, B. Dom, M. Gorkani, J. Hafner, D. Lee, D. Petkovic, D. Stelle, and P. Yanker, “Query by image and video content: The qbic system,” Computer , pp. 23–32, September 1995. http://wwwqbic.almaden.ibm.com. 5. C. Faloutsos, R. Barber, M. Flickner, W. Niblack, D. Petkovic, and W. Equitz, “Efficient and effective querying by image content,” Journal of Intelligent Information Systems (3), pp. 231–262, 1994. 6. R. W. Piccard, A. Pentland, and S. Sclaroff, “Photobook: Content-based manipulation of image databases,” International Journal of Computer Vision 18(3), pp. 233–254, 1996. http://wwwwhite.media.mit.edu/vismod/demos/photobook/. 7. J. R. Bach, C. Fuller, and A. Gupta, “Virage image search engine: an open framework for image management,”
8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27.
28. 29. 30. 31.
32. 33.
Symposium on Electronic Imaging: Science and Technology Storage and Retrieval for Image and Video Databases IV, pp. 76–87, 1996. http://www.virage.com. V. E. Ogle and M. Stonebraker, “Chabot: retrieval from a relational database of images,” Computer , pp. 40–48, september 1995. http://vrw.excalib.com:8015/cst. J. R. Smith and S.-F. Chang, “Visualseek: a fully automated content-based query system,” ACM Multimedia 96 - Boston, MA - USA , pp. 87–98, 1996. http://www.ctr.columbia.edu/visualseek. http://www.excalib.com/products/rw/index.html. http://maya.ece.ucsb.edu/Netra/. J.-Y. Chen, C. A. Bouman, and J. C. Dalton, “Hierarchical browsing and search of large image databases,” IEEE Transactions on Image Processing 9, pp. 442–455, March 2000. A. Grossmann and J. Morlet, “Decompostion of Hardy functions into square integrable wavelets of constant shape,” SIAM J. Math. Anal. 15(4), pp. 723–736, 1984. I. Daubechies, “The wavelet transform, time-frequency localization and signal analysis,” IEEE Transactions on Information Theory 36(5), pp. 961–1005, 1990. S. Mallat, “A theory of multiresolution signal decomposition: the wavelet representation,” IEEE Transactions on Pattern Analysis and Machine Intelligence 11, pp. 674–693, 1989. O. Rioul and M. Vetterli, “Wavelet and signal processing,” IEEE Signal Processing Magazine 8, pp. 14–38, october 1991. Mathsoft wavelets home page http://www.mathsoft.com/wavelets.html. Mathsoft Inc. S. Mallat, “Wavelets for a vision,” Proceedings of the IEEE 84, pp. 604–614, April 1996. C. E. Jacobs, A. Finkelstein, and D. H. Salesin, “Fast multiresolution image querying,” Proceedings of SIGGRAPH95, Los Angeles, California , August 1995. M. K. Mandal, T. Aboulnasr, and S. Panchanathan, “Illumination invariant image indexing using moments and wavelets,” Journal of electronic Imaging 7(2), pp. 282–293, 1998. F. Idris and S. Panchanathan, “Image indexing using wavelet vector quantization,” SPIE 2606, pp. 269–275, 1995. K.-C. Liang and C. C. J. Kuo, “Progressive indexing, retrieval and transmission of wavelet compressed image database,” SPIE 3169, pp. 190–199, 1997. H.-G. Stark, “Image indexing and content base access to databases of medical images with wavelets,” SPIE 2569, pp. 790–800, 1995. ´ Z. Peˇcenovi´c, “Finding rainbows on the internet,” tech. rep., Ecole Polytechnique F´ed´erale de Lausanne, 1998. ´ M. N. Do, “Invariant image retrieval using wavelet maxima moment,” tech. rep., Ecole Polytechnique F´ed´erale de Lausanne, 1998. L. Chen and L. Chien, “Color image segmentation using progressive wavelet transform for image retrieval,” Proceedings of ISIVC’2000 - Rabat - Morocco , pp. 245–252, 2000. J. Landr´e, F. Truchetet, S. Montuire, and B. David, “Content-based multiresolution indexing and retrieval of paleontology images,” SPIE Proceedings of Storage and Retrieval for Media Databases - San Jose - CA - USA 4315, pp. 482–489, 2001. W. Chang, G. Sheikholeslami, J. Wang, and A. Zhang, “Data resource selection in distributed visual information systems,” TKDE 10(6), pp. 926–946, 1998. A. Vailaya, A. Jain, and H. J. Zhang, “On image classification: City images vs. landscapes,” Pattern Recognition 31, pp. 1921–1936, 1998. R. T. Ng and J. Han, “Efficient and effective clustering methods for spatial data mining,” Proceedings of the 20th Very Large Database Conference - Santiago - Chile , pp. 144–155, 1994. T. Zhang, R. Ramakrishnan, and M. Livny, “Birch: An efficient data clustering method for very large databases,” Proceedings of the ACM SIGMOD International Conference on Management of Data - Montreal - Canada , pp. 103–114, 1996. G. Sheikholeslami, S. Chatterjee, and A. Zhang, “Wavecluster: A wavelet based clustering approach for spatial data in very large databases,” VLDB Journal 8(3-4), pp. 289–304, 2000. G. Sheikholeslami and A. Zhang, “Approach to clustering large visual databases using wavelet transform,” Proceedings of the SPIE Conference on Visual Data Exploration and Analysis IV - San Jose - CA - USA 3017, pp. 322–333, 1997.
