Project Detail

Geometry Based Feature Extraction and Analysis for Geo-spatial Datasets

Geometry Based Feature Extraction

Lead: Razdan, Anshuman
Collaborator: Wonka, Peter;
Sponsor: National Geospatial-Intelligence Agency
Date: 07/29/2005  - 07/28/2009

Abstract

This project addresses the problem of creating a general image registration system that 1) enables image to image registration, 2) supports data stemming from different sensors including traditional visible images, synthetic aperture radar (SAR) images, MSI and HSI images, and LIDAR data 3) allows registering data from different sensors and sensors with and without detailed sensor model, and 4) analyzes data for change detection. The system is based on concepts from geometric modeling.

Automated road extraction from an Aerial Image. Useful for comparison and change detection.

An automatic road tracking method is presented for detecting roads from satellite images. This method is based on shape classifcation of a local homogeneous region around a pixel. The local homogeneous region is enclosed by a polygon, called the pixel footprint. We introduce a spoke wheel operator to obtain the pixel footprint and propose a Fourier-based approach to classify footprints for automatic seeding and growing of the road tracker. We experimentally demonstrate that our proposed road tracker can extract the centerlines of roads with sharp turns and intersections effectively, and has relatively small amount of leakage.

First row: input images. Second row: left, result of an area-based method; right, result of our method. To show the visual result for the registration, the second input image is laid on the First one through the transform we got from the corresponding algorithm. Then we calculate the intensity di?erence pixel by pixel. If the difference is equal to zero, we set the corresponding pixel in the result image to 125, which is gray. If the difference is 255, we set it to 255. If the difference is -255, we set it to 0. Other values can be interpolated linearly. If the two images are perfectly registered, then the overlapping part result image should be uniformly gray.

Image registration is a fundamental problem in image processing and often serves as an important preprocessing procedure for many applications such as image fusion, change detection, super resolution, and 3D reconstruction. While we can draw from a large number of solutions, most existing methods are specifically tailored to a particular application. For example, video-tracking focuses on non-rigid transformations with small displacements and medical image registration can take advantages of the uniform background and absence of occlusion. In this paper, we work with remote sensing image pairs that have been geometrically and photometrically rectified to some extent and differ by a global similarity transform. The output of the computation is the set of global parameters of the transform.

In the scope of remote-sensed image registration existing methods fall into two categories: area-based methods and feature-based methods. Area-based methods directly use the whole image or some pre-defined windows without attempting to find salient features. First, a function is defined, such as correlation or mutual information, to evaluate the similarity between the image pairs under a given transformation. The second step is an optimization procedure to find the parameter set that maximizes the function. The drawback of this approach is twofold: searching for a global optimum is computationally expensive and the search space is not smooth due to non-overlapping areas and occlusion in the input images. Therefore, the search process will probably stop at a local optimum instead of a global optimum. In contrast, feature-based methods determine the transformation by matching salient features extracted from the original image pairs. The difficulty of these algorithms is to find corresponding features robustly. If features are matched incorrectly, they become outliers. As features only draw from a smaller local region, the algorithms might create a large number of such outliers and the elimination of outliers is very time consuming and unstable.

Results of the curve-matching algorithm for real aerial images. The first row is the original image pair. The second row is the result of our curve matching algorithm. We show the extracted curves for both images. The computed pairing is indicated by the numbers next to the curves. The third row is the result of the original curve matching algorithm.

We propose a new registration scheme that is a combination of these two approaches: The first step is to segment the input images and to extract closed boundary curves. The second step is to match the extracted curves by an improved curvature scale-space algorithm. In the third step, we try to register the two images using only the regions enclosed by the matched boundaries with an area-based method. This registration framework compares favorably to using feature-based registration or area-based registration alone. Because we use curves as geometric features, our feature-based matching algorithm is more robust than relying on point features. Additionally, the curve based feature-matching algorithm identifies potentially matched regions. This helps the area-based optimization algorithm in the following ways: a) the curve matching gives multiple initial guesses to start a local optimization algorithm b) the search space is smoother compared to the search space induced by the original image pair and therefore there are less chances of getting stuck in a local optimum. c) The computational overhead is greatly reduced. We will demonstrate these advantages through experiments on selected data sets.