A Short Introduction of Multi-Channel CBIR System
(last updated, November 08, 2005)
Multiple Viewpoint
CBIR
How to get multiple viewpoints in CBIR?
How to merge viewpoints?
Example
Future Works
References
Viewpoint is the position or attitude that determines how something is seen, presented, or evaluated (from Merriam-Webster). An information system also provides a viewpoint toward a user's information need. When the user's information need is specific, the retrieval systerm's viewpoint can be affected by queries, retrieval algorithms, document representations, and etc. al, hence multiple viewpoints can be formed. The idea of multi-viewpoint retrieval try to combine (merge) different viewpoints into a better search result.
“Two heads are better than one”. More viewpoints carry more information toward
the user's information need. It is similar to committee members discuss with
each other and then make a decision together, thus reduce any viewpoint's
bias. However, “Truth is not determined by majority vote”. There are works
reported by Beitzel ect. al (Beitzel 2004) that merging different information
sources did not improve performance. Different situation may in favor of different
merging algorithms.
Previous explaination such as Lee's rational for fusion states "different runs retrieve similar sets of relevant documents but retrieve different sets of non-relevant documents" (Lee 1997). However, his coefficients R-OLAP and N-OLAP are neither sufficient conditions nor necessary conditions for performance improvement. Later, Vogt (Vogt 1998) use a linear model to predict performance gain for different runs in TREC5. However, there are too many assumptions in his experimental settings and most are unjustified. It is still unclear in what condition could viewpoints enhance each other to generate a better result. And this question requires furthur research effort.
At least, viewpoints should satisfy two necessary condidtions as:
(1) Viewpoint should provide some “valuable” information. Retrieval system's
performance at least should be better than a random system.
(2) Information is not fully duplicated. There should be partial disagreement
among viewpoints.
Similar ideas have been widely applied in text retrieval and proved to be successful. Several investigators have explored improving retrieval effectiveness by combining multiple information sources, such as different search strategies (fox 1994) and different query representations (belkin 1993). These combining techniques (or referred to as fusion, merging) can result in improved retrieval effectiveness.
With the cheaper storage, faster internet, and off-the-shelf price for personal
digital cameras, image collections expand very quickly today.
Under this background, there are booming needs for image management techniques
and applications, such as image retrieval systems.
CBIR is the application of computer vision to the image retrieval problem, that is, the problem of searching for digital images in large databases. "Content-based" means that the search makes use of the contents of the images themselves, rather than relying on human-inputted metadata such as captions or keywords. A content-based image retrieval system (CBIRS) is a piece of software that implements CBIR (from Wikipedia).
The term of CBIR first appeared in a paper by T. Kato (Kato 92) to describe automatic retrieval of images from a database based on visual features, such as color and shape. The first commercial CBIR system, QBIC (query by image content) was designed by IBM in the early 1990s.
Anything computable from pixels.
At current stage, there is a gap between low-level features of the retrieval system and the high-level semantic concepts of the user, called semantic gap (Gudivada 1995). Compared with text-IR, this problem result in very poor performance of CBIR system.
Figure 1. A query Q is shown entering the CBIR system which is in turn produces a ranked list of result R. We present the viewpoint in a 1-D rank list (referred to as “channel”), sorted in descending order by similarity.
The top 10 images of the channel is listed as
......
Figure 2. Top 10 images of an example CBIR channel.
CBIR system usually employ a query-by-example (QBE) interface. Different example images can be provided as queries. These queries can in turn generate different search results for the same information need.
Figure 3. Two queries for same information need enter the CBIR systems which is in turn produces 2 channels of result R1 and R2.
Figure 4. Suppose the information
need is to retrieve flower. Both a white flower and a pink flower can be query
examples. We can see that they will generate quite different retrieval results
(viewpoints).
Different CBIR systems can generate different search results for the same query.
Figure 5. Two different CBIR systems are used: CBIR1 and CBIR2. A query Q is first transformed into queries for different CBIR systems, respectively, and then enter the CBIR systems which is in turn produces 2 channels of result R1 and R2.
Suppose we use a white flower as the query image to search in two CBIR systems (Basic CBIR and SIMPLIcity), we can get two search results
Figure 6. Search result of Basic CBIR.
Figure 7. Search result of SIMPLIcity.
Figure 8. Four different representations
of images are used: color positive (C+), color-negative (C-), grayscale positive
(B+), and grayscale-negative (B-). A query Q is transformed to four representations,
respectively, and then enter the CBIR system which is in turn produces four
channels of result R1, R2, R3, and R4.
Figure 9. The corresponding four CBIR channels.
