Objectbased Image Representation

1
Object-based Image Representation

• Dr. B.S. Manjunath
• Sitaram Bhagavathy
• Shawn Newsam
• Baris Sumengen
• Vision Research Lab
• University of California, Santa Barbara

2
Outline

• Context and Objective
• Introduction
• Object Extraction and Description
• Time series object coding
• Future research ideas
• Conclusion

3
Context
Large-scale Image Database
Query
user
Retrieved images
Query example Give me all images similar to
image X.
4
Objective

• To develop an object-based image representation
• scheme in order to facilitate the following
• Faster access of data in the context of
  object-based querying
• Reducing required storage space for images
• Relating maps and geographical aerial images
• Study of spatio-temporal relationships in/among
  aerial images
• Note Although our dataset consists of aerial
  images, we expect the scheme to be useful for
  other image datasets as well.

5
The Object-based Approach

• Assumptions
• Useful information (for searching) in images is
  concentrated in smaller regions termed objects.
• Objects are mostly homogeneous in color and
  texture and can be characterized thus.
• Most queries on image databases are in terms of
  objects e.g. Give me all images having a brown
  field.

6
Examples of objects
7
Why Object-based approach?

• Uses semantic information for querying (user
  friendly)
  
  
  
  
• 
  
  
  
  
  
  
  
  
  
  
• Efficient description of images
• We ignore portions of images that would not be
  used for querying
• Potential reduction of storage space
• Store images as collections of objects
• Redundancy removal in time-series of objects

8
Past Research

• Object Extraction Identify and extract semantic
  objects from aerial images.
• So far, we have done this manually
• Working on automatic segmentation using semantic
  models (Sumengen)
• Object Description Find efficient descriptors
  for the objects (Bhagavathy)
• Shape binary alpha plane
• Dominant Colors (Deng, Manjunath)
• Dominant textures

9
object
object
RGB to LUV conversion
24 Gabor filters
24-dim outputs
K-means clustering
K-means clustering
Take means of clusters
Take means and percentages
Convert means to RGB
Dominant texture feature
Dominant color feature
10
Time-series object coding
Objective To apply object-based video coding
(based on MPEG-4) techniques for coding
time-series of objects.
11
Future Research Possibilities

• Storage space issues
• describe image as a collection of objects
• reconstruct image from its objects
• Relation between maps and aerial images
• maps have information, images have data
• Spatial and temporal relationships among objects
• variation of objects with time
• spatial querying (Newsam)
• Application in wireless networks

12
Conclusion

• The object-based approach enables semantic
  querying which is more user-friendly
• Time-series compression of objects reduces
  required storage space for large images
• Potential to reduce required bandwidth for
  wireless transmission of image information
• Enables the study of temporal change in images
• Enables spatial querying