lecture 8

32
Image Indexing and Retrieval Dr Crawford Revie (with thanks to Prof Fabio Crestani) Lecture 8 Multimedia Information Access

Upload: bama-raja-segaran

Post on 26-Jan-2016

213 views

Category:

Documents


0 download

DESCRIPTION

s

TRANSCRIPT

Page 1: Lecture 8

Image Indexing and Retrieval

Dr Crawford Revie(with thanks to Prof Fabio Crestani)

Lecture 8Multimedia Information Access

Page 2: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 2

Outline of lecture

Why do we require image retrieval?Overview of techniquesExamples of systems

NB – if you print these slides out then you will probably want to use colour (or at least for some of them, 18-28 in particular don't make a lot of sense in b/w!)

Page 3: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 3

Motivation and Application Areas

Management of Image ArchivesArt Galleries & MuseumsWWW Image Indexing

Science Databases (Medicine, Astronomy, Geography)Industry specific

Trademark DatabasesTextiles & FabricsAdvertisingArchitecture & Design

Page 4: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 4

Approaches

There are 3 main approaches used in practice:1. Keyword based

manual / semi-automatic / automatic2. Based on visual properties

automatic3. Concept based

mostly manual (still in 'research' mode)

Page 5: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 5

Keyword approach: indexing

Images are annotated using keywords

But:manual annotation is very expensive (as it is exceedingly time consuming)low level visual properties are almost impossible to index consistently using manual mark-upeven for 'high level' properties manual annotation is prone to subjectivity

Take a look at the Google Image Labeler 'game'http://images.google.com/imagelabeler

Page 6: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 6

Keyword approach: retrieval

Since image description is textual, we use an almost straight-forward application of standard IR techniques (stop-word removal, stemming, indexing, etc.)

hypertext links have proved to be useful for retrieving images ("retrieval by browsing")thesauri and vocabularies can be more necessary here than in standard IR (e.g. AAT) – see also later discussion of structured keywords and concept-based retrieval

Page 7: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 7

From manual to automatic keyword assignment

Keywords can be assigned to an image by analysing the text associated with the image

this includes the alt and caption attributes of the <img> and <table> HTML tagstext elsewhere on a web page containing the imagetext of the link pointing to the imageeven the name of the file containing the image

Page 8: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 8

Automatic keywords

Why are thesepictures retrieved?

Page 9: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 9

Automatic keywords

Check the text on the web page; the caption; and the filename

Page 10: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 10

Structured keywords: using a database

Some systemsuse a DBMS to handle keywordsand searches

Page 11: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 11

Visually based approaches

Often referred to as Content Based Image Retrieval (CBIR)

Similarity between query and documents is calculated based on visual features:

colourtextureshape

Visual features may be detected automatically or semi-automatically

Page 12: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 12

Feature vectors

Images are represented using a set of feature vectors

),,( 1 inii IfIfIf K=

Queries are represented with the same set of feature vectors

),,( 1 inii QfQfQf K=

Page 13: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 13

Vector features

Each feature has its own representation range of values, variability, etc.

Feature vectors may provide a “synthetic” view of a certain feature

In IR each word is represented by one feature exactly, but here one image characteristic may be represented by many features

similar to audio retrieval

Page 14: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 14

Similarity functions

Need to choose similarity functions carefully:should be a good approximation to human perceptionof similarity between imagesshould have properties that help speed up computations

Different types of similarity evaluations may need to be combined to compare overall similarity (e.g. shape, colour, texture, etc.)

Page 15: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 15

Colour based retrieval

Arguably easiest; earliest to be usedProcess is as follows:

represent image as a rectangular pixel raster (e.g. 1024 columns and 768 rows)represent each pixel as a quantified colour (e.g. 256 colours ranging from red through violet)count the number of pixels in each colour bin (this will produce a vector representation)compute vector similarity (e.g. using the normalised inner product)

Page 16: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 16

Colour based matching

Let's compare some images retrieved using keyword: Godzilla

Page 17: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 17

Colour histograms (for two samples)

Page 18: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 18

Texture matching

Texture characterizes small-scale regularitycolour describes pixels, texture describes regions

Described by several types of featuressmoothness, periodicity, directionality, etc.

Match region size with image characteristicscomputed using filter banks, Gabor wavelets

Perform weighted vector space matchingusually in combination with a colour histogram

Page 19: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 19

Texture matching

Page 20: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 20

Image segmentation (shape)

Global techniques alone yield low precisioncolour & texture are better at characterising objects, not full images

Segment at colour and texture discontinuitieslike “flood fill” in Photoshop

Represent size shape & orientation of objectse.g. in Berkeley’s Blobworld we use ellipses

Represent relative position of objectse.g. angles between lines joining the centers

Segmentation allows us to perform object rotation and scale-invariant matching

Page 21: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 21

"Flood fill" in Photoshop

More sophisticated techniques are needed

Page 22: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 22

CBIR systems: Examples

Commercial systemsVirageQBIC

AcademicBlobworldVisualSeekChabotViper

Page 23: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 23

QBIC

You can sketchan example of

what you are looking for

Page 24: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 24

QBIC

Can you spotthe similarity to your 'query definition'?

Page 25: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 25

Berkeley Blobworld

Can you spot the similarity to your 'query definition'?

Page 26: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 26

Berkeley Blobworld

Page 27: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 27

Blobworld Segmentation (1)

Page 28: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 28

Blobworld Segmentation (2)

Page 29: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 29

Viper: query by example

Provide examples

by indicatingrelevant images

Page 30: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 30

Viper (QBE)

Retrieves similar

images

Page 31: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 31

Concept based approach

Knowledge of the application domain is requirede.g. indexing of medical images requires knowledge of medicine(!) vocabulary + domain specific

System assigns concepts (index terms) to part of the image:

automatic concept assignment: very imprecise and ambiguous processmanual concept assignment: time-consuming and highly subjective process

Few experiments, semi-automatic seems best so far

Page 32: Lecture 8

MIA Lecture 8 Crawford Revie (2006) 32

Simple image retrieval is commercially availablecolour histograms, texture, limited shape information

Segmentation-based retrieval is still in the lab

Some way off:automatic identification and recognition of objects in images and videosconceptual image retrieval

The future?