Transcript Lecture 5: Multimedia IR: Models and Languages
Special Topics in Computer Science
Advanced Topics in Information Retrieval
Lecture 5
(book chapter 11)
: Multimedia IR: Models and Languages
Alexander Gelbukh
www.Gelbukh.com
Previous Chapter: Conclusions
Inverted files seem to be the best option Other structures are good for specific cases o Genetic databases Sequential searching is an integral part of many indexing-based search techniques o Many methods to improve sequential searching Compression can be integrated with search 2
Previous Chapter: Research topics
Perhaps, new details in integration of compression and search “Linguistic” indexing: allowing linguistic variations o Search in plural or only singular o Search with or without synonyms 3
Motivation
Applications: o office, o CAD, o medical, o Internet Example: o Artists sings a melody and sees all the songs with similar melody 4
What’s different
Different from text IR: o Structure of data is more complex. Efficiency is an issue o Using of metadata o Characteristics of multimedia data o Operations to be performed Aspects: o Data modeling: Extract and maintain the features of objects o Data retrieval: based not only on description but on content 5
Retrieval process
Query specification o fuzzy predicates:
similar to
o content predicates:
images containing an apple
o data type predicates: video, ...
Query processing and optimization o Parsed, compiled, optimized for order of execution o Problem: many data types, different processing for each Answer o Relevance: similarity to query Iteration o Bad quality, so need to refine 6
Modeling
Data modeling
To model is to simplify, in order to make manageable. “We will represent an image as...” o From the user’s point of view o From the system’s point of view (technically) A problem: very large storage size. Modeling needed Objects are represented as feature vectors o Images / Video: shape. House, car, ...
o Sound: style. Music: Merry, sad, ...
Features are defined directly or by comparison o Degree of certainty is stored 8
Multimedia support in commercial DBMSs
(1999) Variable length data. o Non-standard o Different and usually very limited sets of operations SQL3: o provides user-extensible data types o Object-oriented o Implemented partially in many systems Example: data blades of Informix o Content-based functions on text and images o E.g.: date = 1997 AND contains (car) 9
Spatial data types
Informix: 2D, 3D data blades Boxes, vectors, ...
Operations: intersect, contains, center, ...
Text: containWords, ....
Supports query images by content 10
Example: MULTOS
Multimedia document server Documents are described by: o logical structure: title, into, chapter, ... o layout structure: pages, frames, ...
o conceptual structure: allows content-based queries o Docs similar in conceptual structures are grouped into
conceptual types
o Example: Generic_Letter 11
Example of conceptual structure...
12
...continued
13
Image data in MULTOS
Analysis o low level: detect objects and positions o high level: image interpretation Result of analysis: o description of objects found and their classes o certainty values Indices are used for fast access to this info o Object index. Includes pointers to objects and certainty values o Cluster index, with fuzzy clusters of similar images 14
Internet
How Google does it?
No image processing. Textual context!
File names, nearby words Distance from image to words “give me images with
flower
the image” in the file name or near 15
Languages
Query languages
As a query, either a description of the object or an
example object
is submitted o “show me images similar to this one” o in what respects similar?!
Exact match is inadequate. Additional means are needed Content is not a single feature 17
What defines query language
Interface. How to enter the query Types of conditions to specify Handling of uncertainty, proximity, weights 18
Interface
Browsing and navigation Search: description or query by example Query by example: o specify what features are important.
Give me all houses with similar shape but different colors
o Libraries of examples can be provided 19
Conditions...
Attribute predicates o structured content – the predefined types extracted beforehand o Exact match. E.g.: size, type (video, audio, ...) Structural predicates o structure: title, sections, ...
o metadata are used.
Find objects containing an image and a video clip
Semantic predicates o unrestricted content.
o Find all red houses: red = ?, house = ?
Fuzzy
20
... conditions
Predicates o Spatial: contain, intersect, is contained in, is adjacent to ...
o Temporal: Find audio where first politics and then economy is discussed o Spatial and temporal predicates can be combined:
Find clips where the logo disappears and then a graph appears at the same place
A predicate can be applied to a part of document o As path expressions in OO databases 21
Uncertainty, proximity, weights
Similarity function The user can assign importance weights to individual predicates in a complex query This gives ranking, as in text IR The same models can be used, e.g., probabilistic model 22
Examples of query languages: SQL3
Functions and stored procedures: user-defined data manipulation Active database support: database reacts on the events, not only commands. This enforces integrity constraints Good news: rather standard Bad news: no ranking supported!
Effort to integrate SQL3 with IR techniques.
SQL MM Full Text and other similar languages 23
... examples: MULTOS
One of design goals: easy navigation o Paths are supported Identification of components by type, not by position o
All images in the document
, not
the image in 3 rd chapter
Types of predicates: o on data attributes, on textual components, on images (image type, objects contained, ...) Example: 24
MULTOS example
25
Another example of MULTOS
26
Research topics
How similarity function can be defined?
What features of images (video, sound) there are?
How to better specify the importance of individual features? (
Give me similar houses
: similar = size?
color? strructure? Architectural style?) How to determine the objects in an image?
Integration with DBMSs and SQL for fast access and rich semantics o Integration with XML o Ranking: by similarity, taking into account history, profile 27
Conclusions
Basically, images are handled as text described them o Namely, feature vectors (or feature hierarchies) o Context can be used when available to determine features Also, queries by example are common From the point of view of DBMS, integration with IR and multimedia-specific techniques is needed o Object-oriented technology is adequate 28
Thank you!
Till ??, 6 pm
29