learning-based indexing of works of art
DESCRIPTION
Learning-Based Indexing of Works of Art. Kurt Grieb. Presentation Overview. Research Divided into 2 parts Parallel Upgrade of ALIP Structure of Parallelization Results EMPEROR Database Tests Setup of Tests Results. Reasons for Parallelization. - PowerPoint PPT PresentationTRANSCRIPT
Learning-Based Indexing of Works of Art
Kurt Grieb
Presentation Overview
Research Divided into 2 parts Parallel Upgrade of ALIP
– Structure of Parallelization– Results
EMPEROR Database Tests– Setup of Tests– Results
Reasons for Parallelization
ALIP statistical computations are computationally expensive
Corel Image Library Comparison:– 15-20 Minutes– Unacceptable for Web and other applications
Parallelization Concept
One server receives request, divides workload between the total number of clients.
Server
Client1 – 30
Client31 – 60
Client541-570
Client571-600
. . . .
Parallelization Structure
PERLGUI CLIENTS
Server
Request With URL Range of Concepts
Likelihoods Best Fit
Results
The Speedup of ALIP
y = 0.6275x + 0.6884
0
1
2
3
4
5
6
0 2 4 6 8 10
Number of Processors Used
Sp
eed
up
Series1
Linear (Series1)
Results
600 concepts can now be computed in roughly 40 seconds over 30 processors.
Roughly ideal speedup More processors on a smaller size reduces
efficiency of speedup
The EMPEROR Library
1700 Images Chinese Historical Images
The Testing
2 sets of tests (9 and 20 concepts) 4 runs per set (best, worst, 2 random) 4 sizes per run (3, 6, 9, 12)
Set 1
Best Sub Worst Sub Random 1 Random 2
Size 3 Size 6 Size 9 Size 12
Set 2
Best Sub Worst Sub Random 1 Random 2
Size 3 Size 6 Size 9 Size 12
Motivation For Test Structure
Effects of more specific classes Effects of different training classes Determine reasonable training sizes
Results
Set 1 Total Percentages
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
3 6 9 12
Sample Size
% C
orr
ect
Worst Case
Random 2
Random 1
Best Case
Random Generation
Interesting Cases / Notable Trends
Set One vs. Set Two The Black and White Sketches General Trends vs. Specific Classes Weak Classes Misclassification of Similar Objects
– Black & White Images vs. Text – All faces vs. Color/BW Faces– Faces and Upper Bodies
The Black and White Sketches
Performed the best of all classes Accuracies of 99% over all tests Due to difference between this class and
most other classes
Interesting Cases / Notable Trends
The overall accuracy of all classes went up with more training
In certain classes, the accuracy went down as all concepts were trained with more imaging
Paintings Accuracy
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
3 6 9 12
Number of Training Images
Per
cen
tag
e co
rrec
t
Paintings
Weak Classes
In certain concepts a weak class outperformed other classes
Could be due to openness of concept spaces
Horses Comparison
0.2
0.3
0.4
0.5
0.6
0.7
0.8
2 4 6 8 10 12
Training Image Size
Per
cen
tag
e C
orr
ect
Best Case
Worst Case
Random 1
Random 2
Misclassification of Similar Objects
Pictures with more than one concept in them sometimes can confuse ALIP
Misclassification of Similar Objects
Further Work
Overlapping of Concepts 3-D representations of objects Improved Accuracy of ALIP Current Results are Promising
ABSTRACT
Digital images are widely and readily in use. Text based indexing of these images is becoming tougher as the number of digital images grows. Therefore, Content Based Image Retrieval is becoming a more viable alternative because of the ability to automate this process. Dr. Wang’s Automatic Linguistic Indexing of Pictures shows great promise as a Contend Based Image Retrieval system. Our lab is looking to expand this indexing of pictures for artistic/historical purposes, which are harder to classify due certain characterizes of these pictures. Additionally, some upgrades need to be made to ALIP in order to convert it to a more user-friendly, mainstream program. I present the results of the upgrades to ALIP and the experiments conducted on a historic image database.