Synopsis Of Current Image Search Systems - MULTIMEDIA

Some other current image search engines are mentioned here, along with URLs for each (more URLs and resources are in the Further Explorations section). The following is by no means a complete synopsis. Most of these engines are experimental, but all those included here are interesting in some way. Several include query features different from those outlined for C-BIRD.

intra - frame and Inter - frame video locales algorithm results: (a) original images; (b) intra - frame results; (c) inter - frame results

intra - frame and Inter - frame video locales algorithm results:


One interesting feature in QBIC is the metric it uses for color histogram difference. Instead of simple histogram Intersection, the metric recognizes that colors that are similar, such as red and orange, should not have a zero intersection. Instead, a color - distance matrix A is introduced, with elements


Here, dij is defined as a three - dimensional color difference (using Euclidean distance, or any other likely distance — sum of absolute values, say). Then a histogram - difference D2 is defined as follows [20]:

difference D2

Vector z is a histogram - difference vector (for vectorized histograms). For example, the histogram - difference vectors z would be of length 256 if our two - dimensional chromaticity histograms were 16 x 16.

The striking feature of this metric is that it allows us to use simple differences of average three - dimensional color as a first screen, because the simpler metric is guaranteed to be a bound on the more complex one in above equation.

QBIC has been developed further since its initial version and now forms an essential (and licensable) part of IBM's suite of Digital Library products. These aim at providing a complete media - collection management system.

An interesting development in the QBIC research effort at IBM is the attempt to include grayscale imagery in its domain, a difficult retrieval task. QBIC can combine other at­tributes with color - only - based searches — these can be textual annotations, such as captions, and texture. Texture, particularly, helps in gray level image retrieval, since to some extent it captures structural information in an image. Database issues begin to dominate once the data set becomes very large, with careful control on cluster sizes and representatives for a tree - based indexing scheme.

UC Santa Barbara Search Engines

Alexandria Digital Library (ADL) is a seasoned image search engine devised at the University of California, Santa Barbara. The ADL is presently concerned with geographical data: "spatial data on the web". The user can interact with a map and zoom into a map, then retrieve images as a query result type that pertain to the selected map area. This approach mitigates the fact that terabytes, perhaps, of data need to be stored for LANDS AT images, say. Instead, ADL uses a multiresolution approach that allows fast browsing by making use of image thumbnails. Multiresolution images means that it is possible to select a certain region within an image and zoom in on it.

  • NETRAis also part of the Alexandria Digital Library project. Now in its second generation as NETRA II, it emphasizes color image segmentation for object­or region - based search.

  • Perception - Based Image Retrieval (PBTR)aims at a better version of learning and relevance feedback techniques with learning algorithms that try to get at the underlying query behind the user's choices in zeroing in on the right target.

Berkeley Digital Library Project

Text queries are supported, with search aimed at a particular commercial or other set of stock photos. The experimental version tries to include semantic information from text as a clue for image search.


Chabot is an earlier system, also from UC Berkeley, that aims to include 500,000 digitized multiresolution images. Chabot uses the relational database management system POSTGRES to access these images and associated textual data. The system stores both text and color histogram data. Instead of color percentages, a "mostly red" type of simple query is acceptable.


Blobworld was also developed at UC Berkeley. It attempts to capture the idea of objects by segmenting images into regions. To achieve a good segmentation, an expec­tation maximization (EM) algorithm derives the maximum likelihood for a good cluster­ing in the feature space. Blobworld allows for both textual and content - based searching. The system has some degree of feedback, in that it displays the internal representation of the submitted image and the query results, so the user can better guide the algorithm.

Columbia University Image Seekers

A team at Columbia University has developed the following search engines:

  • Content - Based Visual Query (CBVQ),developed by the ADVENT project at Columbia University, is the first of the series. (ADVENT stands for All Digital Video Encoding, Networking and Transmission.) It uses content - based image retrieval based on color, texture, and color composition.

  • VisualSEEk is a color - photograph retrieval system. Queries are by color layout, or by an image instance, such as the URL of a seed image, or by instances of prior matches. VisualSEEk supports queries based on the spatial relationships of visual features,

  • SaFe, an integrated spatial and feature image system, extracts regions from an image and compares the spatial arrangements of regions,

  • WebSEEk collects images (and text) from the web. The emphasis is on making a searchable catalogue with such topics as animals, architecture, art, astronomy, cats, and so on. Relevance feedback is provided in the form of thumbnail images and motion icons. For video, a good form of feedback is also inclusion of small, short video sequences as animated GIF files.


The Informedia Digital Video Library project at Carnegie Mellon University is now in its second generation, known as Informedia n. This centers on "video mining" and is funded by a consortium of government and corporate sponsors.


MetaSEEk is a meta - search engine, also developed at Columbia but under the auspices of their IMKA Intelligent Multimedia Knowledge Application Project. The idea is to query several other online image search engines, rank their performance for different classes of visual queries, and use them selectively for any particular search.

Photobook and FourEyes

Photobook was one of the earlier CBIR systems developed by the MIT Media Labora­tory. It searches for three different types of image content (faces, 2 - D shapes, and texture images) using three mechanisms. For the first two types, it creates an eigenfunction space - a set of "eigenimages". Then new images are described in terms of their coordinates in this basis. For textures, an image is treated as a sum of three orthogonal components in a decomposition denoted as Wold features.

With relevance feedback added, Photobook became FourEyes. Not only does this system assign positive and negative weight changes for images, but given a similar query to one it has seen before, it can search faster than previously.


MARS (Multimedia Analysis and Retrieval System) was developed at the University of Illinois at Urbana - Champaign. The idea was to create a dynamic system of feature representations that could adapt to different applications and different users. Relevance feedback, with changes of weightings directed by the user, is the main tool used.


Visual Information Retrieval (Virage) operates on objects within images. Image in­dexing is performed after several preprocessing operations, such as smoothing and contrast enhancement. The details of the feature vector are proprietary; however, it is known that the computation of each feature is made by not one but several methods, with a com­posite feature vector composed of the concatenation of these individual computations.


Visual Information Processing for Enhanced Retrieval (VIPER) is an experimental system that concentrates on a user - guided shaping of finer and finer search constraints. This is referred to as relevance feedback. The system is developed by researchers at the University of Geneva. VIPER makes use of a huge set of approximately 80,000 potential image features, based on color and textures at different scales and in a hierarchical decomposition of the image at different scales. VIPER is distributed under the auspices of the open software distribution system GNU ("Gnu's Not Unix") under a General Public License.

isual RetrievalWare

Visual RetrievalWare is an image search technology owned by Convera, Inc. It is built on techniques created, for use by various government agencies for searching databases of standards documents. Its image version powers Yahoo's Image Surfer. Honeywell has licensed this technology as well. Honeywell x - rayed over one million of its products and plans to be able to index and search a database of these x - ray images. The features this software uses are color content, shape content, texture content, brightness structure, color structure, and aspect ratio.

All rights reserved © 2018 Wisdom IT Services India Pvt. Ltd Protection Status