Geometry Augmented SURF with Modified Sobel for Improved Affine Invariance in Image Matching

Abstract

Reliable image matching and alignment is a key issue in difference extraction of aerial images. This paper presents an affine, scale and rotation-invariant method for aligning images taken at different timelines. SURF feature points index pair polling is used to detect best candidate match from an image library against a reference image. SURF is used to ensure speedy match detection as a large library is being scanned. The two images are then coarse-aligned using a statistical model. A modified sobel operator is used to ensure complete edge detection along six orientations. Since SURF is not satisfied for affine invariance, a geometry based approach is used to discard undesired differences. The resulting difference helps locating new structures/ buildings. This integrated approach allows difference extraction in affine environments while satisfying robustness and low computational complexity. The results show upto 90% increase in correlation after alignment between the reference and matched image. The augmented approach increased the probability of detecting valid differences while suppressing the false detections upto 99%.