Patch-based image classification using image epitomes
The goal of this project is to design a classifier capable of learning
to recognize whether an image belongs to a certain class or not.
Example classes would be "human face", "beach", "tiger", etc.
The classifier would be given a series of training images and
told which are positive and negative examples for the classification.
From these training examples, the classifier would extract common
features that could be used to decide whether a new image belongs to
the class or not.
This project attempts to acheive this by finding patches that are
common to all or most of the positive example images, but absent from
all or most of the negative example images. These patches could
be called "discriminative patches" because they discriminate between
positive and negative examples.
In order to find these discriminative patches, this project uses a
probabilistic image representation known as an "image epitome".
This representation condenses an image into a smaller collection
of patches that can be quilted to reconstructed the original
image.
The basic method employed by this project is to
- Construct a training collage of positive and negative example images
- Extract the epitome of this collage
- Try to find patches in the epitome that map preferentially to the positive example images
- Use those patches as discriminative patches for classifying new images
Presentation (ppt)
Final report (pdf, html)
Experimental results
References
Image Epitomes
N. Jojic, B. J. Frey, and A. Kannan, Epitomic analysis of appearance and shape, Proc. 9th Int. Conf. Computer Vision, 2003. (webpage)
Video Epitomes
V. Cheung, B. J. Frey, and N. Jojic, Video epitomes, Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2005. (webpage)
Object Recognition
R.
Fergus, P. Perona, A. Zisserman, Object Class Recognition by
Unsupervised Scale-Invariant Learning, Proc. of the IEEE Conf on
Computer Vision and Pattern Recognition, 2003.