Hough-SIFT: Robust Image Registration for Linear Structures via Hough Space

Standard image registration techniques, such as the Scale-Invariant Feature Transform (SIFT), often fail in environments dominated by repetitive linear structures like shutters or louvers. In these scenes, local keypoint descriptors become ambiguous, leading to frequent registration failures. This paper investigates whether transforming images into Hough space—where linear structures are localized into distinct peaks—can restore the discriminative power of SIFT and enable robust registration.

The author introduces Hough-SIFT, a three-step registration pipeline. First, the input image is processed using a novel 'texture-preserving' Hough transform, which maps linear structures into Hough space while retaining enough local texture information for feature description. Second, standard SIFT keypoints are extracted and matched within this Hough domain. Finally, the homography between images is estimated by minimizing a custom cost function that accounts for the dual relationship between points and lines in projective geometry, using RANSAC for robust outlier rejection.

Experiments on both synthetic datasets and real-world videos demonstrate that Hough-SIFT significantly outperforms conventional SIFT in scenes with strong linear structures. While SIFT frequently fails in these challenging environments, Hough-SIFT maintains high success rates and low photometric errors. In 'normal' scenes, Hough-SIFT performs comparably to standard SIFT, confirming that the method is a robust, general-purpose alternative for applications like electronic image stabilization (EIS) where linear structures are common.

Registration failure in video processing can lead to severe visual artifacts, such as ghosting or unnatural jitter. As camera technology advances—particularly with higher zoom capabilities that emphasize linear patterns—the need for robust registration becomes critical. Hough-SIFT provides a training-free, computationally efficient solution that integrates seamlessly into existing computer vision pipelines without requiring large datasets or complex deep learning models.