We have proposed two methods of modeling object classes using texture motifs:

1. Modeling with texture motifs only
2. Modeling with texture motifs and their spatial relationships

1. A texture motif is a texture that is common to and characteristic of a class of objects. Examples below.

2. Premise: Often objects of interest are created by geographic processes that result in distinctive image textures. For example, an instance of the "harbor" class is shown below.

3. Object Modeling: Given a set of example objects from a class, we learn the texture motifs of the class using a Gaussian mixture model (GMM). The features are outputs of Gabor filters at 5 scales and 6 orientations, arranged as 30-D vectors.

4. Motif Identification: Now, given any object from the class, we can identify and label its texture motifs. Classification of an unlabeled object can also be done using this model.

5. Examples: Red = moored boats, Dark blue = water

Harbor object		Labeled motifs

Harbor object		Labeled motifs

1. The spatial arrangement of texture motifs may be an additional characteristic of an object class. For example, both parks and golf courses have trees and grass but they look different.

Consider a sequence of pixel-blocks (left image) in the object, each described with a texture feature. Hidden Markov models (HMM) are used to learn the texture motifs and simple spatial relationships among them. States Si (texture motifs) are hidden, only the texture features are observable. Inter-state transition probabilities aij model inter-motif spatial adjacency relationships. Combination of 1-D HMMs are used for capturing 2-D information.

2. Armed with models for M classes, we can classify a new object into one of these.

The table shows the average (over a test set) "unlikelihood" that a model generated a test object. We call this "unlikelihood" because the number is a negated log-likelihood, i.e. the lower this number, the higher the likelihood.

S. Bhagavathy, S. Newsam, and B.S. Manjunath, "Modeling Object Classes in Aerial Images Using Texture Motifs," International Conference on Pattern Recognition, Quebec City, August 11-15, 2002. Abstract

S. Newsam, S. Bhagavathy, and B.S. Manjunath, "Modeling Object Classes in Aerial Images Using Hidden Markov Models," International Conference on Image Processing, Rochester, September 22-25, 2002. Abstract

S. Newsam, L. Wang, S. Bhagavathy, and B. S. Manjunath, "Using texture to analyze and manage large collections of remote sensed image and video data," Journal of Applied Optics: Information Processing, vol. 46, no. 2, Jan 2004. Abstract