Shape Modeling for Remote Sensing Image Segmentation
 

This project uses statistical learning for object shape modeling. The approach uses the concept of sign distance maps in building a probabilistic representation for the object’ shape. The statistical model of the shape is used in many applications, especially in segmentation. Classification of remote sensing data is used to illustrate the application of this project.

 
Research Team
Aly A. Farag CVIP Director farag@cvip.uofl.edu
Refaat M. Mohamed Research Assistant refaat@cvip.uofl.edu
Ayman S. El-Baz Research Assistant elbaz@cvip.uofl.edu
Methods

This project aims to incorporate the shape constraints in image segmentation. The shape priors are modeled by generating a signed Distance Map (SD-Map) is generated for each class in the data set. This SD-Map measures the relative positions of pixels in the reference shape image with the shape boundary. The SD-Map is probabilistic in nature which allows for statistical modeling of the object shape. The MF-SVM algorithm is used for estimation the shape pdf. Bayesian classification based on using the shape pdf as class priors is used for the image segmentation.

Results

RGB from a 6-band multispectral data

 Classified Image

Water class points

Signed distance map of the Water class

pdf of the Water class model

Transportation class points

Signed distance map of the Transportation  class

pdf of the Transportation  class model

Publications

Refaat M. Mohamed, Ayman S El-Baz, and Aly A. Farag, “Shape Constraints for Accurate Image Segmentation with Applications in Remote Sensing Data,” Submitted to the eighth International Conference on Information Fusion, Philadelphia, PA, USA , July, 2005.

Acknowledgement / Sponsors

We would like to thank the Air Force Office for Scientific Research