Sub-pixel Image Registration Using 2D Registration Methods

Image registration serves as a fundamental technique in computer vision, with its core objective being the spatial alignment of images captured under different conditions of the same scene. When sub-pixel registration accuracy is required, traditional integer-pixel displacement methods become insufficient. Achieving sub-pixel registration typically involves three critical phases.

The initial phase employs phase correlation for coarse registration. This method leverages the translation property of Fourier transform by calculating the cross-power spectrum between two images to estimate overall displacement, effectively mitigating the impact of illumination variations. After obtaining the initial displacement estimate in the frequency domain, the system proceeds to the fine registration stage. In code implementation, this typically involves using FFT (Fast Fourier Transform) functions and calculating the normalized cross-power spectrum to identify peak coordinates for displacement estimation.

The fine registration stage commonly utilizes upsampling interpolation strategies. By performing zero-padding upsampling in the frequency domain or employing interpolation methods like cubic splines in the spatial domain, the image resolution is virtually enhanced. When combined with optimization algorithms such as least squares, displacement measurements surpassing single-pixel accuracy can be calculated. Particularly when using optimization functions incorporating phase information, sub-pixel accuracy in displacement estimation can be significantly improved. Programmatically, this involves implementing interpolation algorithms and optimization functions that minimize residual errors between transformed and target images.

The final verification phase is achieved through resampling techniques. Using bilinear or bicubic interpolation, the reference image undergoes sub-pixel displacement resampling, followed by similarity measurement with the target image. Common evaluation metrics include mutual information and normalized cross-correlation. When the similarity reaches a predetermined threshold, the registration process is considered complete. This approach holds significant application value in medical imaging, remote sensing image processing, and related fields. Code implementation typically requires creating custom functions for similarity measurement and threshold comparison to validate registration quality.