a lightweight image retrieval system for paintings t. lombardi, s. cha, and c. tappert january 19th,...

23
A Lightweight Image A Lightweight Image Retrieval System Retrieval System for Paintings for Paintings T. Lombardi, S. Cha, and C. T. Lombardi, S. Cha, and C. Tappert Tappert January 19th, 2005 January 19th, 2005

Upload: oswald-lambert

Post on 26-Dec-2015

223 views

Category:

Documents


5 download

TRANSCRIPT

Page 1: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

A Lightweight A Lightweight Image Retrieval Image Retrieval

System for System for PaintingsPaintings

T. Lombardi, S. Cha, and C. T. Lombardi, S. Cha, and C. Tappert Tappert

January 19th, 2005January 19th, 2005

Page 2: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

IntroductionIntroductionStudents of art history learnStudents of art history learn

three primary skills:three primary skills:

Formal analysisFormal analysis ComparisonComparison ClassificationClassification

How can computer science How can computer science

contribute to the contribute to the developmentdevelopment

of these skills?of these skills? Figure 1: Girl with a Pearl Earring, Jan Vermeer, 1665

Page 3: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Working HypothesisWorking Hypothesis

An Interactive Indexing and Image An Interactive Indexing and Image Retrieval System (IIR) for fine-art Retrieval System (IIR) for fine-art paintings can aid students in these paintings can aid students in these endeavors by providing:endeavors by providing: a mathematical summarization of an imagea mathematical summarization of an image a measurable basis for comparing two a measurable basis for comparing two

imagesimages an elementary way to classify an image an elementary way to classify an image

relative to those in a database relative to those in a database

Page 4: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Previous WorkPrevious WorkWe synthesize the goals of two research areas:We synthesize the goals of two research areas:

Classification of paintings:Classification of paintings: R. Sablatnig, P. Kammerer, and E. Zolda, “Hierarchical Classification of

Paintings Using Face- and Brush Stroke Models”, in Proc. of the 14th International Conference on Pattern Recognition (1998).

D. Keren, “Painter Identification Using Local Features and Naïve Bayes”, in Proc. of the 16th International Conference on Pattern Recognition (2002).

Image retrieval which aims to bridge the semantic gap:Image retrieval which aims to bridge the semantic gap: J. Corridoni, A. Del Bimbo, and P. Pala, “Retrieval of Paintings using Effects

Induced by Color Features”, in Proc. of the International Workshop on Content-Based Access of Image and Video Databases (1998).

Can we construct a feature set that satisfies the objectives of both areas while providing analytically relevant data to students?

Page 5: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

System OverviewSystem Overview

The system consists of two major The system consists of two major components:components:

Image Database Image Database stores images, thumbnail images, and stores images, thumbnail images, and

extracted features for later retrieval extracted features for later retrieval and analysis.and analysis.

Graphical User Interface Graphical User Interface provides interactive query capabilities provides interactive query capabilities

to the end userto the end user

Page 6: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Database ConstructionDatabase Construction

An XML index file stores extracted An XML index file stores extracted features and control information.features and control information.

A file system stores images and A file system stores images and thumbnail images.thumbnail images.

The open design of the database The open design of the database contributes to the goals of ease of contributes to the goals of ease of use and exchange of information.use and exchange of information.

Page 7: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Database Construction – Database Construction – Cont.Cont.

Figure 2: XML Index File Figure 3: File System

Page 8: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Global Feature Global Feature ExtractionExtraction

Two different kinds of features are extracted:Two different kinds of features are extracted: Palette features Palette features

concern the set of colors in an image (color concern the set of colors in an image (color map)map)

examples: palette scopeexamples: palette scope Canvas features Canvas features

concern the spatial and frequency distribution concern the spatial and frequency distribution of colors in an image (image index)of colors in an image (image index)

examples: max, min, median, mean (for each examples: max, min, median, mean (for each color channel)color channel)

Page 9: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Sample Feature Set Sample Feature Set

Feature NameFeature Name Description and NotesDescription and Notes

MaxMax Max value of H, S, and V channelsMax value of H, S, and V channels

MinMin Min value of H, S, and V channelsMin value of H, S, and V channels

MeanMean Mean of H, S, and V channelsMean of H, S, and V channels

MedianMedian Median of H, S, and V channelsMedian of H, S, and V channels

Standard Dev.Standard Dev. Std of H, S, and V channelsStd of H, S, and V channels

Color EntropyColor Entropy Measures the frequency distribution of colorMeasures the frequency distribution of color

Line CountLine Count Normalized number of detected edges – Sobel Normalized number of detected edges – Sobel edge detectoredge detector

Intensity MeanIntensity Mean Arithmetic mean of values in a grayscale imageArithmetic mean of values in a grayscale image

Table 1: Sample Features used for Web Museum Interactive Test

Page 10: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Example: Palette ScopeExample: Palette Scope

Palette ScopePalette Scope -- the total number of unique colors -- the total number of unique colors used in an image.used in an image.

We expect Dali’s piece to have a higher palette We expect Dali’s piece to have a higher palette depth than Mondrian’s work.depth than Mondrian’s work.

Figure 4: Hallucinogenic ToreadorSalvador Dali, 1970

Figure 5: Composition with Large Blue Plane,Red, Black, Yellow, and GrayPiet Mondrian, 1921

Page 11: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Example: Palette Scope – Example: Palette Scope – Cont.Cont.

Formal definition of Palette Scope (U):Formal definition of Palette Scope (U):

U = C/PU = C/P

WhereWhereC=Total # of unique colors measured C=Total # of unique colors measured

in RGB or HSV triples.in RGB or HSV triples.P= Total # of pixels in an image.P= Total # of pixels in an image.

Page 12: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Example: Palette Scope – Example: Palette Scope – Cont.Cont.

ArtistArtist Total Pixels Total Pixels (P)(P)

Total Colors Total Colors (C)(C)

Palette Depth Palette Depth (U)(U)

MondrianMondrian 359700359700 22422242 0.006230.00623

DaliDali 165775165775 38993899 0.023510.02351

Table 2: Palette Scope statistics.

We see that Dali uses more of the color spectrum than Mondrian.

Palette depth is an important feature for artist and period style identification because many styles are defined by color, i.e. Picasso’s Blue Period and fauvism.

Page 13: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Graphical User InterfaceGraphical User Interface

The GUI consists of three primary The GUI consists of three primary windows for:windows for: AnalysisAnalysis ComparisonComparison ClassificationClassification

Page 14: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Analysis WindowAnalysis Window

Figure 6: The Analysis Window

Page 15: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Comparison WindowComparison Window

Figure 7: The Comparison Window

Page 16: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Classification WindowClassification Window

Figure 8: The Classification Window

Page 17: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Test ResultsTest Results

Two types of tests were conducted:Two types of tests were conducted:

Feature testsFeature tests Feature tests focus on the accuracy of Feature tests focus on the accuracy of

specific collections of features.specific collections of features. Interactive testsInteractive tests

Interactive tests assess the accuracy Interactive tests assess the accuracy of the system as a whole.of the system as a whole.

Page 18: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Feature TestFeature Test

Training SetTraining Set Test SetTest Set Percent CorrectPercent Correct

3636 3636 9494

200200 200200 8888

200200 200200 8383

Figure 9: Les Demoiselles d’Avignon,Pablo Picasso, 1907.

Figure 10: Road with Cypress and Star,Vincent Van Gogh, 1890.

Table 3: Feature test to distinguish the work of Picasso and Van Gogh.

Page 19: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Initial Interactive TestInitial Interactive Test

Database of 10 works of each of the following Database of 10 works of each of the following ten artists: ten artists:

Braque, Cezanne, De Chirico, El Greco, Braque, Cezanne, De Chirico, El Greco, Gauguin, Gauguin,

Modigliani, Mondrian, Picasso, Rembrandt, Modigliani, Mondrian, Picasso, Rembrandt, and Van and Van

Gogh.Gogh.

Training SetTraining Set Testing SetTesting Set Percent Percent CorrectCorrect

100100 9090 8181

Table 4: Initial Interactive Test

Page 20: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Interactive Test: Web Museum

ArtistArtist Training Training SetSet

QueriesQueries SuccessSuccess PercentPercent

AertsenAertsen 99 99 55 55.655.6

El GrecoEl Greco 1010 77 44 57.157.1

HopperHopper 1010 77 33 42.942.9

MalevichMalevich 1010 1111 88 72.772.7

MonetMonet 1010 1010 66 60.060.0

MorisotMorisot 1010 1111 77 63.663.6

RembrandtRembrandt 1010 3333 2525 75.875.8

RenoirRenoir 1010 3838 1414 36.836.8

TurnerTurner 1010 1010 44 40.040.0

VelazquezVelazquez 1010 88 88 100.0100.0

OverallOverall 500500 293293 165165 56.356.3

Table 5: Results from Web Museum Interactive Test

Page 21: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

EvaluationEvaluation ofWeb Museum Test ResultsTest Results

Overall result: 56.3% accuracy Overall result: 56.3% accuracy 36.3% better than blind guessing (10 36.3% better than blind guessing (10

guesses/50 artists = 20%)guesses/50 artists = 20%) Dissecting the classification Dissecting the classification

mistakes reveals some intelligent mistakes reveals some intelligent mistakesmistakes Rembrandt is most often confused with Rembrandt is most often confused with

Caravaggio, Ast, and VermeerCaravaggio, Ast, and Vermeer

Page 22: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

ConclusionsConclusions Simple palette and canvas features are Simple palette and canvas features are

sufficient for an interactive classification sufficient for an interactive classification systemsystem

A single feature set can serve for A single feature set can serve for classification and image retrieval classification and image retrieval applicationsapplications

A general feature set can adequately serve A general feature set can adequately serve for educational applicationsfor educational applications

Although showing promise, we currently Although showing promise, we currently have a low confidence system have a low confidence system

Page 23: A Lightweight Image Retrieval System for Paintings T. Lombardi, S. Cha, and C. Tappert January 19th, 2005 January 19th, 2005

Electronic Imaging 2005

Future WorkFuture Work

Add texture featuresAdd texture features Improved color features: hue Improved color features: hue

histogramshistograms Improved distance metrics: modulo Improved distance metrics: modulo

comparison of hue histogramscomparison of hue histograms Test against larger datasetsTest against larger datasets