INEX 2006

1
INEX 2006

• Supervisors
• Associate Professor Dr. Shlomo Geva
• Dr. Dian Tjondronegoro
• Dr. Jinglan Zhang
• Presenters
• CherHan Lau, Tavitt Chaiyawan

2
Introduction

• Internet is growing rapidly with faster
  connection at reasonable price
• Price drop in physical storage
• Increment in types of multimedia documents on
  Internet

3
Motivation

• Current image search system lack of functionality
• Based on keyword search on Filename or Text
  around the image
• It is efficient, but human subjectivity and word
  ambiguity is the main problem to overcome

4
Word ambiguity
5
Human perspective subjectivity
Sunset? Boat? Or Beach?
6

• Use traditional Textual database to describe the
  image is unpractical and consume large volume of
  database.
• For Eg. All the following information store in
  Database

Where is the car now? At the city or downtown?
The colour of the car
The style of the car hatch back or sedan
The number of the lights
7
CBIR Structure
Semantic level features Face recognition, Shape
recognition
High level features Face detection, shape
detection
Medium level features Hough Transform, Texture
Low level features RGB Histogram, HSV Histogram,
Edge Detection
8
System Framework
9
Features used

• Hough transform
• Texture (Object and Background)
• RGB Histogram
• YCbCr Histogram
• HSV Histogram

10
Combination

• The system allows various feature to be combined
  with different weights.
• The main combination used in the submission in
  Hough ObjectRGB and Texture (H.O.T)

11
Text and Image Fusion
1000 1001 1002 1003 1004 1005 1006

12
CBIR Search Results
13
RGB Histogram
14
Hough transform
15
Texture
16
H.O.T Combined feature
17
Text and CBIR fusion results
18
Text only
19
Text and CBIR Fusion
20
CBIR only
21
Text and CBIR fusion
22
INEX Evaluation results
23
Text only
24
Text Image Features
25
Implementation
26
CBIR(Content Based Image Searching using Java GUI
)

• Application overview and demo

27
Standalone Tools and Platform

• Eclipse Java IDE
• JFrame Builder incompatible with Eclipse GUI
  Builder.
• Java NetBeans visual database structure view
  for Apache Derby Database
• MTLAB for Image extraction
• WinXP

28
Web-based Development Tools

• Eclipse WTP IDE
• Java/JSP tools
• Compiling and Debugging
• Apache Derby GUI Database Explorer
• MyEclipse Workbench
• Deploy code from Eclipse to Tomcat
• Apache Tomcat Server
• Web application server

29
Problems during design

• GUI Design with limited function Visual Tools
• Apache Derby Database Keyword Object
• Efficiency of searching speed separate table
• Java Memory Boundary Xmx, Xms
• Java Hash Table Size
• Microsoft Access DB Boundary
• Program Initialization Memory Problem
• Java this.dispose()

30
Problem during implementation

• Java memory overrun
• Apache Tomcat memory overrun
• Calculation method during multiple features
• Image path can only be under Apache Tomcat
  directory

31
Working Diagram
Feature Extraction by MATLAB RGB RGB
Background RGB Object Full Texture
Over 100000 images
32
Reflection and Conclusions

• Not suitable for Rapid Development and deployment
  Java Programming Language
• Image Searching based on the matching features
  still hard to provide the relevant results
• Too much data amount for this project and
  application

33
Standalone to Web-baseImplementation
34
Why Implement Web-base?

• Make system more worthwhile
• Everyone can benefit from it
• Back-link to Wikipedia website
• Non-Platform restriction
• Can run on any device that have
• web browser and internet connection
• (PDA (tested) or mobile phone)

35
Calculation Performance
36
Web-base Prototype
37
Future Work

• Integrated with Dr. Shlomos Textual Retrieval
  System (currently semi-automated)
• Classification and Clustering
• Principal Component Analysis
• Weighage between Text Search and CBIR
• Image to text search

38
Future work Image to text search
Trinidad Tobago
This flag belongs to which country ?