Joint Shape and Appearance Features for Registration

Members: Azubuike Okorie, Sokratis Makrogiannis

Introduction

Image registration is the process of aligning two images of the same scene for analysis.

Image registration can be used in target recognition, for matching of targets with real-time image of a scene, in remote sensing for monitoring climatic changes and land use using satellite images, in biomedical imaging for matching images from two modalities for diagnosis, monitoring growth of tumor/cancer using CT-scans, and in computer vision and pattern recognition-for matching stereo images for object/shape recognition.

Purpose and Goal

Many image registration methods have be proposed in literature, such as Speeded-Up Robust Features (SURF) [1] , Binary Robust Invariant Scalable Keypoints (BRISK) [2], HARRIS corner detectors [3], and Minimum Eigenvalue methods. Efficient methods rely on local-invariant image features [4] for registration. In this project, our goal is to create features from image regions whose efficiency would be measurable to existing methods.

Our algorithm is designed based the major steps for feature-based registration, which are feature detection, feature extraction, feature matching and estimation of geometric transform.

Method

Feature detection: The pair of images are first segmented into regions using watershed segmentation and then the regions are represented by their centroids.

Feature Extraction. We extract a combined region features composed of normalized shape properties, intensity features and histogram of region intensities (HORI).

Feature Matching: We find correspondences based using a joint cost function which gives us a weighted sum of two Euclidean distances for shape and intensity features, and histogram intersection for HORI feature. Strong and unambiguous matches are extracted based on predefined thresholds.

Estimation of Geometric transform: We use maximum likelihood estimation sampling consensus (MLESAC) for estimation of Geometric transform.

Validation

Our result are validated by calculating the root-mean-squared-error (RMSE) between estimated geometric transform and the ground truth transform described by methods described in [5]

Result:

The illustration below is an example of image registration of a satellite image based on our approach. The RMSE for the process is 0.462, which infers that our method is capable of registering remote sensing images up to subpixel level.

 

References

[1]

Bay, Herbert,; Tinne , Tuytelaars; Luc, Gool Van;, "Surf: Speeded up robust features," European conference on computer vision. Springer, Berlin, Heidelberg, , 2006.

[2]

S. Leutenegger, C. Margarita and S. Y. Roland , "BRISK: Binary robust invariant scalable keypoints.," Computer Vision (ICCV), 2011 IEEE International Conference on. IEEE, 2011.

[3]

Harris, Chris; Mike , Stephens;, "A combined corner and edge detector," Alvey vision conference, vol. 15, no. 50, 1988.

[4]

Tuytelaars, Tinne; Krystian , Mikolajczyk;, "Local invariant feature detectors: a survey," Foundations and trends® in computer graphics and vision, vol. 3, no. 3, pp. 177-280, 2008.

[5]

A. Goshtasby, "Image registration by local approximation methods.," Image and Vision Computing, vol. 6, no. 4, pp. 255-261, 1988.