Journal of Visual Languages and Computing (1998) 9, 103—117
REGULAR ARTICLE
A Multidimensional Image Browser L. CINQUE,* S. LEVIALDI,* A. MALIZIA
AND
K. A. OLSENst
*Pictorial Computing Laboratory, Dip. Scienze dell’Informazione, Universita’ di Roma, Via Salaria 113, 001895 Roma, Italy, Cinque.dsi.uniroma1.it,
[email protected] sMolde College, Britv. 2, N-6400 Molde, Norway,
[email protected] Submitted 8 October 1996; accepted 5 November 1997 We present a browsing tool for content-based image retrieval. Images are retrieved from the database based on both textual and geometrical attributes. The resulting image collection is presented in a user-defined multidimensional visual information space, which acts as an interface to the underlying image database.
( 1998 Academic Press Limited
Keywords: browsing, content-based image retrieval, visualization
1. Introduction THE ADVANCE of pictorial databases, collections of photos, clip art files, etc., makes it interesting to consider retrieval tools that work directly on the images. Such tools could be used instead of, or in combination with, traditional text retrieval tools working on image annotations. One such tool is model-based vision. Here knowledge about the object structures is utilized in the recognition process. The process consists of two tasks. First, a structural description of the visual query, for example a sketch, is provided by the user. This description is then used in the second part, matching the description to images in the database. To make the matching more efficient, structural descriptions of all the database images will usually be created in a preprocessing phase. That is, in order to formalize the process of image searching, we perform the process on a level where both queries and database contents may be formalized. As an example, we consider the user that requires a picture of an airplane, seen in the air. An example of what the user needs is shown in Figure 1. Such a request to the database may be formalized by searching for the keyword ‘airplane’ in a caption or picture annotation record, perhaps trying to combine this with other words such as ‘in the air’ or ‘flying’. However, the annotation may be lacking, the keywords provided by the user may not be present in the annotation, or the user may not be able to provide appropriate keywords. As an alternative, or in combination to text-based retrieval, the user may try to search directly on the image geometrical attributes. By using the t Corresponding author.
1045-926X/98/010103#15 $25.00/0/vl970066
( 1998 Academic Press Limited
104
L. CINQUE ET AL.
Figure 1. Boeing 737-600 in the air
Figure 2. Airplane model
model-based vision approach, this may be done by creating a sketch of the required image. An example is shown in Figure 2. This model is then compared to models of all images in the database. Common to both textual and model-based retrieval is the fact that the actual retrieval process is performed on a lower formalization level than the original request. The user wants something similar to the image shown in Figure 1, but has to formalize the request on a lower level, as the keyword ‘airplane’ or as the model in Figure 2. While the user’s request is on a semantic level, we process the query on a lexical, or at best, on a syntactic level. This gap between ‘information need’ and request will manifest itself in the retrieval results. Since the query given to the image database is formalized at a lower level than the original request, there will be relevant objects in the database that are not retrieved and irrelevant objects that are retrieved. This is in sharp contrast to what we expect for other types of database systems, where queries may be formalized at a higher level. For example, a bank clerk who asks the system for a printout of all accounts with a negative balance will consider anything apart from a 100% effectiveness an error, either in the data or in the system itself. Thus, as long as there is a gap between request and formalized query, we may not expect full retrieval efficiency in image databases. If we want to find all relevant objects, this may only be achieved if the user performs the retrieval manually, scanning all the objects in the database. The size of most image databases makes this approach impractical. A natural solution is then to combine these two methods. First, a simplified automatic search process that retrieves a subset of the database is performed. Then the user interacts directly with this subset to find the relevant images. Most bibliographic search systems follow this principle. The query returns a list of objects, usually allowing the user to retrieve detailed information on every object, e.g. heading, abstract or full text. The list of objects may be presented chronologically, in alphabetic order or may be sorted according to some priority factor. For example, keywords that are found in
A MULTIDIMENSIONAL IMAGE BROWSER
105
headings may give higher priority than keywords found in the main text. Alternatively, the number of occurrences of each keyword may determine the priority. Similar methods are used in image-retrieval systems, where a set of images is returned, often presented in a miniaturized (iconized) form [1—3]. Most systems present retrieval results as a single list of objects. We feel that this one-dimensional presentation is not in accordance with how most people tackle hard problems. In situations where it is difficult or impossible to get the 100% correct answer, we often try different approaches towards a solution. For example, let us return to our request for an image of an airplane and let us assume that we have a system that may perform searches based on textual annotations and geometric attributes. In order to find the right picture, we may try different searches, using both textual and geometric queries. Each query produces a list of images. Each list is presented as a separate entity, in isolation from the results from the other queries. If the lists are short, we may take a closer look at every object. With a large database or broad queries, the lists may be long, making it impractical to perform a manual check on each object. What we need is a holistic approach, where the results of all the queries may be presented in the same information space. This will provide additional information, making it easier to determine which objects are relevant. In this paper, we describe a multidimensional presentation tool for image retrieval. The tool, ImageVIBE (visualization browsing environment), allows the user to view the retrieved objects in a multidimensional information space, created by the user. For example, the object returned from our airplane example may be presented in an information space defined by keywords such as airplane and flying, together with geometrical attributes, such as model similarity, orientation and color. Since we assume that readers are familiar with traditional text-based retrieval techniques, we shall start by introducing model-based retrieval.
2. Model-based Retrieval In a number of applications, the goal of a vision system is to identify and locate a specific object in the scene. In such cases, a vision system must have a full knowledge of the shape of the desired object. This knowledge is provided through a ‘model’ of the object, typically, a model includes information on the shape, texture and context of such an object in a scene. A system that makes use of an object model is referred to as a ‘model-based’ vision system, and the general problem of identifying the desired object is referred to as object recognition. For image retrieval, the recognition phase may be initiated by the construction of a sketch of the required image. This is then transformed to a structural model, which is compared to similar formalizations of the database images. This process often involves two search levels. The outer level is the search through the database to select those candidates that are most likely to match the model. Then, to find the best image an inner-level search is required. This is the process of evaluating how ‘closely’ a model corresponds to the extracted structure. The problem of selecting the structures that are the model components is directly related to the problem of model definition. However, the shape of objects is generally the most significant feature for object recognition, location and description. In turn,
106
L. CINQUE ET AL.
shape may be composed of many different features like edge, corner and boundary curvature like concavity, convexity, compactness, symmetry, elongation, etc., to provide accurate and practical representations of an object class. These features represent specific higher-level primitives that correspond to physically meaningful properties of the objects, being less sensitive to intensity variations than the original noisy gray-level values. Usually, the decision of which features to use is rather subjective and application-specific. Moreover, the images to be retrieved must be relevant to the users’ goal. Previous approaches to the image-retrieval problem have been in one of two directions. In the first direction [4, 5], image contents were described as image attributes; attributes are extracted manually from the images and are managed within the framework of conventional database systems. The second approach [6—8] emphasizes the importance of an object recognition system as an integral part of the image-retrieval system to overcome the limitations of manually attribute-based retrieval. However, object recognition is a computationally expensive task and makes the approach unsuitable even for moderately sized image databases. Furthermore, image-retrieval systems based on this approach tend to be domain-specific. Recent research [9, 10] focuses on bridging the gap between the previous two approaches to the problem. The primary emphasis has been on developing domainindependent image-retrieval systems that provide the ability to exploit their contents without the need to perform the object recognition task at query processing time. These efforts have culminated in the introduction of novel image representations and of image data descriptions [11, 12]. Most image retrieval work has concentrated on techniques for matching either (a) whole images or (b) features as shape, color, texture and spatial and/or temporal relationships extracted from images. Several systems that attempt to provide a general retrieval capability have been developed. In such systems, the pioneering work was done by Chang [13], who developed an approach to the design of an integrated database system that store tabular, graphical and image data. The main thrust of the work was to represent pictorial information by both physical and logical pictures. The physical picture is the original picture, and the logical picture is a hierarchically structured collection of picture objects. In this system, the retrieval of images is achieved through attribute matching, spatial relation matching, structural matching and similarity matching using various similarity distances. Recently, color [14], texture [15] and spatial arrangements [16, 17 ] have also become important features to be used in the classical task of image retrieval in a pictorial database. One of the most interesting general systems is QBIC [18]. This general-purpose system allows retrieval of images by color, texture and the shape of image objects or regions. Retrieval by color is based on color histogram matching, retrieval by texture uses coarseness, contrast and directionality features, while shape-matching is based on classical shape measures such as area, circularity, eccentricity, major axis orientation and algebraic moment invariants. Another database system that uses general techniques is ART MUSEUM [19, 20]. The system includes a visual interface where the user can enter a hand-drawn sketch or a full color image of a painting in order to retrieve matched images from the database. The matching process is not performed on full images, but on pictorial indices. These
A MULTIDIMENSIONAL IMAGE BROWSER
107
are constructed from the original images by a process of normalization, gradient edge detection, thinning and shrinking. A set of interactive tools for browsing and searching images was developed in the system called Photobook [9]. This system uses several different semantic-preserving representations for images and provides the user with retrieval tools based on these representations. Photo-book allows queries by appearance, shape and texture. Appearance refers to the technique of matching with eigenimages; shape refers to the object form appearing on the image. The technique used, the finite-element method shape representations for deformable shapes is presented by Sclaroff and Pentland [21]. Finally, a new model for texture matching based on the Wold decomposition for regular stationary stochastic processes is applied [22]. Most of the above techniques are global in that they compare some general feature, or features of an image, like contour signature, histograms, eigenimages, etc. Consequently, all of the above approaches either apply a matching procedure to every image in the database or maintain a single external index for image retrieval. Our approach capitalizes on previous work [23, 24], exploiting a number of features that provide indices to be matched: (i) boundary shape, (ii) contour signature, (iii) curve distance and (iv) color. Moreover, these features are also perceptually significant to humans.
3. Image VIBE A complementary activity to image retrieval is image browsing. Recently, a new semantically based screen layout methodology, visualization browsing environment ( VIBE) has been introduced in order to perform browsing in a human perceptual way [25]. The methodology allows for the presentation of data objects in a multidimensional, user-defined information space. The idea behind this methodology is to present an intuitive mapping between data objects and their position in the information space. This information space is defined through a set of points of interest (POIs). Each POI represents a property of the data objects. The data objects are positioned in this space, according to the score on each POI. An example of a VIBE diagram is shown in Figure 3. This example shows an information space with four data objects. The space is constructed by positioning three POIs on the display: A, B and C. Each POI is represented by a circular icon. The POIs
Figure 3. ImageVIBE diagram (example)
108
L. CINQUE ET AL.
represent user-selected properties of the data objects, and ‘influence’ these objects to a degree determined by a similarity score; the data objects are represented by rectangular icons. From the position of the rectangular icons, we see that only one object is influenced by all three POIs. This is found in the middle of the diagram. There is one icon on top of B, indicating a score on this POI but a zero score on the two others. Another icon is found between A and B, indicating an object that obtained an even score on these two POIs and a zero score on C. Similarly, the object close to C has an A and C score, and a zero B score. However, due to the proximity to C, the influence from C must be higher than from A. While the icon position gives information on relative POI scores, the size of the object icon gives an indication of the maximum score. Thus, a large icon will indicate an object that received a high score on at least one of the POIs, presumably the proximate POI. The actual position (xd , yd ) of an object icon d in a visualization space with n POIs is determined by the simple formula n
(xd , yd )" + adi (ui , vi ) i"1
where adi is the normalized score ( + adi"1 for all POIs i ) of object d on POIi , (ui , vi ) being the position of this POI. Scores are also normalized over all data objects m for all POIs i, thus, m
∀i + adi"1. d"1
Normalization of scores allows us to combine POI scores of different types in the same display, e.g. POI scores based on geometrical and textual attributes. The positioning scheme used by VIBE guarantees unambiguous positioning in a three POI information space. With higher dimensions there may exist more than one score combination for an icon position. However, VIBE offers the user a set of tools to explore the information space. The user may click on an icon to obtain all available data on this icon, e.g. the image itself. POI scores can be visualized by a line drawn from the icon towards the POIs, where the length of this line indicates score value. By adding a color to a POI, all icons influenced by this POI will be drawn in this color. Alternatively, a color may be added to a category attribute of the object, presenting all objects of this category in this color. The information space is dynamic in the sense that new POIs may be added to the diagram, existing POIs may be removed or repositioned. The system will then automatically reposition all icons in the display. Based on the methodology from VIBE a special system, ImageVIBE, has been developed (the system is written in Visual Basic 5.0 and runs on a Windows 95 platform). ImageVIBE is specially designed for visualization of image databases, and includes a set of content-based image-matching algorithms. The system expects a database of images as input, each image described as a file in a standard graphics format. In addition, an annotation file may be given for each data object (a text file). In principle, the system may also work on composite objects of both graphics and text, e.g. a document from a modern word processor.
A MULTIDIMENSIONAL IMAGE BROWSER
109
4. ImageVIBE Browsing A snapshot of the ImageVIBE prototype user interface is shown in Figure 4. In this example, the information space is defined by the POIs passenger-airplane, horizontal, vertical, model-1, blueBackground and redObject. From the figure, we see that the user has clicked twice on an icon, thus asking the system to open the appropriate object (an airplane image). The ImageVIBE display gives an intuitive overview of the data collection. For example, if we are mainly interested in images of passenger-airplanes similar to model-1 with a horizontal orientation, we should examine images in the left part of the diagram. If the color and vertical orientation of the image are of no interest, we may eliminate these POIs. We then get a simplified information space, defined by passenger-airplane, model-1 and horizontal only. The result is shown in Figure 5. Here we only consider scores from these three POIs. Some of the images displayed have only non-zero scores on one or two POIs. For example, all icons on model-1 to horizontal line have a zero score on passenger-airplane. Further, a set of icons has fallen on top of the passenger-airplane POI, indicating a zero score on the other two active POIs. Note that an underlying line replaces an icon falling on top of another. If we want to study the influence on the redObject POI (giving a score to images with a red foreground color) we may reintroduce this POI. If the displacement function is on, ImageVIBE will indicate the movement of each icon by drawing a line from the former to the new position. From Figure 6 we see clearly the data objects influenced by the redObject POI.
Figure 4. Snapshot of ImageVIBE user interface
110
L. CINQUE ET AL.
Figure 5. Simplified ImageVIBE diagram (example)
Figure 6. Displacement function (example)
A MULTIDIMENSIONAL IMAGE BROWSER
111
Through these and other functions the user is offered a wide range of tools to manipulate the set of data objects, presented as an icon collection in the display. The icon to data object connection is handled by ImageVIBE, and the user will always be able to retrieve a data object by ‘clicking on’ the appropriate icon (as previously seen from Figure 4). The contents of the image and the annotation file (if available) will then be presented. Since we are working on images it seems of interest to be able to display these directly, i.e. by letting ‘postage-stamp’ versions of the images replace our rather boring rectangle icons. While such diagrams certainly would allow for nice demonstrations, we are afraid that this approach would not scale up to real problems. The idea of a visualization browser is to give the users an overview over large data collections. Even miniaturized versions of each image would be so pixel-intensive that only a limited amount of images could be displayed at the same time. Even with an initial filter query, one often has collections of several hundred images. Thus, a pixel miserly icon is needed to avoid a cluttered diagram. We have tested applications with up to 25 POIs and 4000 data objects, and even with our simple icons, the displays become cluttered. However, even with such large collections it is possible to identify clusters and exceptions, and to get an overview of the data, since the icons are so simple. Still image data are best presented as images. Within the constraints of the above discussion, an image-cursor has been implemented. By moving this cursor over the icons, a simplified model (outline only) of each image will be presented (see also Section 6).
5. Image POIs ImageVIBE recognizes the POI classes keyword, model, orientation, color and representational properties. Standard scoring functions are implemented for these classes. In addition, the user may provide customized scoring functions. Based on these classes, the user may define any number of POI objects; for example, as previously shown in Figure 4 (horizontal POI, vertical POI, etc.). Each of these different classes is discussed in some detail below.
5.1. Keyword-based POI Class This POI class is defined as a set of keywords. The score is computed as the sum of the occurrences of the keywords found in the image object annotation. For example, the passenger-airplane POI of the previous examples was defined through the keywords, ‘Airbus 320’, ‘Boeing 737’, ‘MD80’, etc., as seen from the example in Figure 7. Added weight may be given to keywords used in annotation headings. In order to speed up the scoring calculations, ImageVIBE will perform a preprocessing of the annotations, creating a frequency table of all words (excepting stop-words) of each annotation. Scoring is then performed by retrieving the actual frequencies of the keywords from this table. Some image collections, e.g. a clipart library, may organize the images into different categories. ImageVIBE may use this information directly, i.e. not as a POI, but by letting the user assign a color code to each category. Object icons will then be colored
112
L. CINQUE ET AL.
Figure 7. Specification of a keyword-based POI (example)
according to category. For example, we may give a separate color code to categories such as supersonic, jet and propeller airplanes in a database of airplanes.
5.2. Model-based POI Class With a model-based POI, the score is computed based on the similarity between a user-defined sketch and the database images (actually a comparison to model descriptions of the images). Each sketch may be used to describe up to six different POIs. Figure 8 shows an example where four POIs are specified. In a preprocessing phase, ImageVIBE will find the contour of all images in the database, using a standard eight-connectedness border detection algorithm [26] after the images have been transformed to black and white. The scoring functions below use the contours extracted by this process. The minimum enclosing rectangle defines the aspect ratio of the required image. It is based on the sketch, which in the simplest case may be a rectangle, and compared to the aspect ratio of the models of the database images. The aspect ratio is found by dividing the difference between the maximum and minimum y- and x-values of an image, after the main axis of an image has been aligned with a coordinate axis. This scoring function is invariant with regard to translation, scaling and rotation. The signature, as seen in the example in Figure 9, is based on sampling the distance from the center of an object to the contour at discrete angles, where the object center C is
A
n
xi ; i"1 n
C" +
n
B
yi . i"1 n +
The distances are converted to a grammar of distance types (based on length), reducing the comparison to a pattern-matching problem. Thus, the scoring function is translation, scaling and rotation-invariant. In addition, the sum of the distances for the images is used directly as score for the signature distance POI.
A MULTIDIMENSIONAL IMAGE BROWSER
113
Figure 8. Specification of a set of model-based POIs (example)
Figure 9. An image and its signature curve
Figure 10. An image and its distance curve
The distance curve of the concavities of an image is a function of the distance between the contour of the image and its convex hull computed for all points along the convex hull. The convex hull of an arbitrary set of point S is defined as the smallest convex polygon containing S. An example of an object and its distance curve is given in Figure 10. As with the signature, the distance curve comparison is performed by pattern matching. In addition, the height of the maximum peak of the images is used to determine the score for the peak height POI. Orientation scores are computed as the cosine between the orientational vector specified by the user (the arrow in Figure 8) and the major axis of the image.
114
L. CINQUE ET AL.
5.3. Color-based POI Class A color-based POI is defined by selecting a color, or a color range, from a color scale. Color definitions may also be specified together with the sketch (model), defining color attributes for different parts of the object. For color matching of a colored sketch to an image, we use the histogram intersection method described by Swain and Ballard [27]. A color histogram is obtained by discretizing the color space into n bins, and then computing the color histogram of the model M and the image I. The similarity score S is then defined as the intersection of these two histograms: n
S" + min (Ij , Mj ) j"1
This scoring function is translation and rotation invariant. S is also robust with regard to scaling, as a result of the minimum function.
5.4. POIs Based on Representational Properties POIs can be specified based on representational properties, such as: z z z z
image size (in bytes) number of pixels number of different colors color depth.
These values may be mapped directly to score values for the corresponding POIs.
5.5. User-defined POI Scoring Functions ImageVIBE accepts a scoring database in a standard format as alternative input, giving a score for each image for each POI. The prototype works on a Microsoft Access database. Each data set is defined as a table, each POI as a column and with an additional column for the file reference and category. Each row represents an individual image.
6. Discussion Content-based image retrieval is an ‘open’ application, in the sense that it is often difficult for the user to formalize a request. Thus, it is important that the user is allowed to use all available data in the retrieval process, i.e. annotations, shape, orientation, etc. The main advantage of the ImageVIBE visualization methodology is that results on each of these dimensions are integrated in the same display. This gives the user an overview of the image collection, which may be used as a basis for further pursuit of the right image. In this way, ImageVIBE extends the notion of a database view to encompass the complete object collection. ImageVIBE is a highly dynamic system. The retrieval and analyzing phases are performed in visual space, where the user may change the display through direct manipulation techniques. The display will be updated immediately, since scores and
A MULTIDIMENSIONAL IMAGE BROWSER
115
image references are available in computer memory. However, when a new POI is specified, or an existing POI changed, it will be necessary to compute new score values for these POIs on all images. This process is also performed interactively. During a preprocessing phase, ImageVIBE will have parameterized all images based on the standard POI classes. Thus, the actual scoring functions are performed in parameter space, as lookups in frequency tables, peak-type grammar matching, color histogram intersection, etc., all of which may be performed within seconds on a modern PC. ImageVIBE does not recognize any strong dividing line between query and result. Just as a sketch (query) may be viewed as a preliminary image (result), the VIBE display is used both to form queries and to evaluate results. We feel that this closed feedback loop is especially important in image browsing, where results strongly influence successive queries, even the initial ‘information need’. This lack of distinction between query and result can also be seen from the example in Figure 11. This display presents a sample of a data collection, a user-specified number of image models which belong to a database of fish images. These are presented in a display of five POIs: minimum enclosing rectangle (MER), signature, distance measurement curve, peak heights and orientation. Note that only the images with a maximum score are visualized as models, using the standard icon for the others. This display may be viewed as a map of the information space, more to give the user an overview of the visualization space than to be used for image retrieval (it would be overcrowded if all the images were presented). The map may give answers to questions as to which POIs are the most relevant; where are the interesting areas of the display; how do the images change along this axis? Thus, the map will give a visual and intuitive explanation of how the different scoring functions perform on the current image collection. Remembering this map, or keeping it in a separate window, the user may require ImageVIBE to present the full image collection, now with the standard rectangular icons.
Figure 11. Map of the information space (example)
116
L. CINQUE ET AL.
7. Conclusion and Future Work A browser for image retrieval has been presented. This is based on a multidimensional visualization system (ImageVIBE), which allows the user to define a visualization space consisting of POIs (points of interest). Each POI represents a property of the images, model similarity, color, orientation, representational properties or annotations. The influence of a POI on an image is represented by a numerical score, computed by a set of predefined scoring functions or provided by the user. Since POI scores are normalized, different types of POIs may be used in the same display, giving the user a multidimensional overview of the image collection, where each image is represented by a simple rectangular icon. The user may manipulate this display through a set of direct manipulation visualization tools, allowing an efficient cognitive feedback loop. This display acts as an interface between the user and the underlying image collection. A prototype of ImageVIBE has been developed. This includes scoring functions for keyword-, model-, orientation- and color-based retrieval. We are currently implementing additional retrieval functions. A central task will be to perform an empirical evaluation of the system on a large image database.
References 1. H. K. T. Wong & I. Kuo (1982) GUIDE graphical user interface for database exploration. In: Proceedings of the VIII VLDB Conference, Mexico City, pp. 22–31. 2. D. Fogg (1984) Lessons from a living in a database graphical query interface. In: Proceedings of HCM/SIGMOD Conference on Management of Data, Boston, MA, pp. 100–106. 3. J. A. Larson (1986) A visual approach to browsing in a database environment. IEEE Computer 19, pp. 62–71. 4. W. Grosky & R. Mehrotra (guest eds) (1989) Special issue on image database man-agement. Computer 22, 12. 5. G. Jung & V. Gudivada (1994) Adaptive query reformulation in attribute-based image retrieval. In: Proceedings of the 3rd Golden West International Conference on Intelligent Systems. Kluwer Academic Publishers, Boston, MA, pp. 673–774. 6. V. N. Gudivara & V. V. Raghavan (1995) Introduction content-based image retrieval systems. Computer 28, 9, 18–22. 7. M. Flickner, H. Sawhney, W. Niblack, J. Asley, Q. Huang, B. Dom, M. Gorkani, J. Hafner, D. Lee, D. Petkovic, D. Steele & P. Yanker (1995) Query by image and video content the QBIC system. Computer 28, 9, 23–32. 8. V. N. Gudivara, V. V. Raghavan & K. Vanapipat (1996) A unified approach to data modelling and retrieval for a class of image database applications. In: Multimedia Database Systems Issues on Research Directions (S. Jajodia & V. Subrahmanian, eds), Springer. 9. A. Pentland, R. W. Picard and S. Sclaroff (1994) Photobook tools for content-based manipulation of image databases. In: Proceedings of Storage and Retrieval for Image and Video Database II, Vol. 2, 185. SPIE, Bellingham, Washington, pp. 34–47. 10. A. Del Bimbo & P. Pala (1996) Effective image retrieval using deformable templates. In: 13th ICPR, Vienna, pp. 120–124. 11. T. M. Strat & M. A. Fischler (1991) Content-based vision recognizing objects using information from both 2-D and 3-D imagery. IEEE Transactions on PAMI 13, 1050–1065. 12. R. Mehrotra and J. E. Gary (1995) Similar-shape retrieval in shape data management. Computer 28, 9, 57–62. 13. S. K. Chang, J. Reuss & B. H. McCormick (1978) Design considerations of pictorial database system. International Journal on Policy Analysis and Information Systems 1, 2.
A MULTIDIMENSIONAL IMAGE BROWSER
117
14. A. Del Bimbo, M. Mugnaini, P. Pala, F. Turco & L.Verzucoli (1997) Image Retrieval by color regions. ICIAP, Florence, pp. 180–187. 15. V. E. Ogle & M. Stonebraker (1995) CHABOT retrieval from a relational database of images, special issue on content-based image retrieval systems. Computer 28, 49—56. 16. V. N. Gudivada & V. V. Raghavan (1995) Design and evaluation of algorithms for image retrieval by spatial similarity. ACM Transactions on Information Systems 13, 115–144. 17. A. Del Bimbo, E. Vicaro & D. Zingone (1994) Sequence retrieval by contents through spatio temporal indexing. IEEE Symposium on Visual Languages, pp. 216–223. 18. W. Niblack (1993) The QBIC project querying images by content using color, texture, and shape. IBM Report RJ 9203. 19. K. Hirata & T. Kato (1992) Query by visual example. In: Advances in Database Technology. Springer, Berlin, pp. 56–71. 20. T. Kato, T. Kurita, N. Otsu and K. Hirata (1992) A sketch retrieval method for full color image database. In: Proceedings of the 11th International Conference on Pattern Recognition, pp. 530—533. 21. S. Sclaroff & A. Pentland (1994) Object recognition and recognition using modal matching. In: Proceedings of the 2nd CAD-Based Vision Workshop, pp. 258–265. 22. R. W. Picard & F. Liu (1994) A new wold ordering for image similarity. In: Proceedings of the Conference on Acoustic Speech and Signal Processing. 23. M. De Marsico, L. Cinque & S. Levialdi (1997) Indexing pictorial document by their content. A survey of current techniques. Image and Vision Computing 15, 119–141. 24. G. C. Peritore (1995) Basi dati pictoriali Un sistema interattivo con interrogazione per analogia. Thesis, Rome University. 25. K. A. Olsen, R. R. Korfhage, K. M. Sochats, M. B. Spring & J. G. Williams (1993) Visualization of a document collection the VIBE system. Information Processing & Man-agement 29, 69–82. 26. A. Rosenfeld & A. C. Kak (1982) Digital Picture Processing, 2nd edn. Computer Science and Applied Mathematics, Vol. 2, pp. 220–223. 27. M. J. Swain & D. H. Ballard (1991) Color indexing. International Journal of Computer Vision 7, 11–32.
