Using multi-scale density for local feature-based registration in SAR imagery

We present a registration algorithm for automatic, robust SAR (Synthetic Aperture Radar) image alignment. The registration problem is handled using sparse feature representation, which comprises local feature localization and description. The feature location is determined by detecting bifurcation structure in edge image and its orientation is assigned using corresponding bifurcation structure type. Then, the local structure is characterized by distinctive non-parametric cross-scale descriptor derived from image patches which extracted in every level of multi-scale pyramid and centered at relevant bifurcation structure. We adopt an elaborate feature mismatches identification strategy to perform Generalized Hough Transform and robust RANSAC fitting in sequence. Accordingly, a group of more accurate warp parameters can be derived even when outliers are predominant in primary feature match set. This approach provides robust matching across a substantial rang of distortion and is less sensitive to speckle noise as well as lack of stable details in SAR imagery. In experimental results, we demonstrated the effectiveness of this approach for natural SAR images.