Region Based Image Annotation Through MultipleInstance Learning

1
Region Based Image Annotation Through
Multiple-Instance Learning

• By Changbo Yang
• Wayne State University
• Department of Computer Science

2
Annotation by Region
An image contains several regions, and each
region may have different contents. The
incomplete information provided by the training
images. The annotation information for a training
image is usually available at the concept level
(annotation to the image ), but NOT at the
content level (annotation to region) .
A large number of irrelevant noisy regions, exist
in the training set for keyword tiger.
3
Bayesian Framework for image annotation

• Consider image annotation as a problem of image
  classification, in which each keyword is treated
  as a distinct class label.
• Given the feature vector of a testing image I,
  what is the probability that I belongs to class
  w?
• One of the key steps in Bayesian classification
  is to select the most representative image region
  for a keyword .
• a large amount of irrelevant noisy regions in
  training images. MIL could be used to predict the
  representative region

4
What MIL can do?

• Multiple-instance learning a variation of
  supervised learning.
• MIL Problem Each bag may contain many instances.
  A bag is labeled positive even if only one of the
  instances is positive. A bag is labeled negative
  only if all the instances in it are negative.
• In region based image annotation, each region is
  an instance, and the set of regions that comes
  from the same image can be treated as a bag.
• One way to solve MIL problem is to examine the
  distribution of these instances, and look for an
  instance that is close to all the instances in
  the positive bags and far from those from
  negative bags.
• MIL algorithm Diverse Density, MIL-SVM, ASVM and
  so on.

5
Region based image annotation
Predict the most possible representative region
of a semantic meaning by relevant and irrelevant
images.