Analysis of image registration noise due to rotationally dependent aliasing

Harold S. Stone, Bo Tao, Morgan McGuireJVCIR 2003 · DOI

TL;DR

This paper identifies and quantifies "rotationally dependent aliasing" as a significant source of noise in image registration using phase correlation. By developing and testing techniques like windowing and frequency exclusion, they demonstrate substantial improvements in registration precision and signal-to-noise ratio, achieving significantly higher peak correlations than previously reported.

What Problem Does This Paper Address?

Image registration, the process of aligning two similar images, is crucial for many applications. Phase correlation is a powerful technique for this, especially when dealing with translations, rotations, and scale changes. However, the precision of phase correlation can be degraded by noise. This paper focuses on a specific type of noise: "rotationally dependent aliasing." This arises because the discrete Fourier transform (DFT) of a sampled image does not commute with rotation, unlike its continuous counterpart. This non-commutativity introduces artifacts that can lead to inaccurate registration, particularly in the presence of sharp image features. The authors aim to understand, quantify, and mitigate these aliasing-induced errors to improve the robustness and accuracy of phase correlation.

Methods

The paper delves into the mathematical underpinnings of Fourier transforms and image rotations, first establishing that commutativity holds for infinite continuous images. It then systematically analyzes the breakdown of this commutativity for finite continuous and, crucially, finite discrete images, highlighting the role of interpolation and aliasing.

Key techniques and contributions include:

  • Mathematical Analysis: Derivations showing how rotationally dependent aliasing arises in the discrete Fourier domain, leading to different aliasing effects depending on the order of rotation and aliasing operations.
  • Windowing: Demonstrating the effectiveness of applying windows (specifically, a Blackman window) to the image before transformation. This smooths sharp boundaries, reducing high-frequency energy that exacerbates aliasing.
  • Frequency Exclusion: Investigating the impact of removing low-frequency components from the Fourier transform. They show that excluding frequencies within a radius of N/4 of the origin significantly reduces aliasing effects.
  • Phase Correlation Variants: Exploring techniques like performing phase correlation over half the Fourier plane, which more than doubles the signal-to-noise ratio.
  • Experimental Validation: Conducting a series of experiments to quantify the impact of various noise sources (interpolation, aliasing, coordinate conversion) on both normalized correlation and phase correlation.

Datasets

The paper uses a single example image: an aerial photograph of an agricultural scene. This image is cropped into a circular region for experiments. No new benchmarks are introduced; the focus is on analyzing the behavior of existing registration techniques on this representative image.

Key Insights and Ideas

  • DFT Non-Commutativity with Rotation is Fundamental: The core insight is that for discrete images, the Fourier transform and rotation operations do not commute. This isn't just an artifact of interpolation; it's a fundamental property of discrete sampling that leads to rotationally dependent aliasing.
  • Aliasing is Rotationally Dependent: Unlike other noise sources, the aliasing artifacts introduced by rotation change depending on the specific orientation of the image. This makes it a particularly tricky problem for rotation-invariant registration.
  • Sharp Edges are the Culprits: High-frequency components, often generated by sharp image boundaries, are the primary drivers of severe rotationally dependent aliasing.
  • Windowing is Crucial: Applying a window function (like Blackman) effectively smooths these sharp edges, drastically reducing high-frequency content and, consequently, aliasing.
  • Low-Frequency Exclusion is Effective: Removing a central region of the Fourier domain (e.g., within N/4) is a powerful way to suppress the aliasing artifacts that tend to manifest in these low-frequency areas.
  • Phase Correlation is Sensitive, but High SNR: While phase correlation is more susceptible to aliasing noise than normalized correlation, it can achieve exceptionally high signal-to-noise ratios when properly filtered.

Key Results

  • Significant SNR Improvement: The combination of techniques (Blackman windowing, N/4 low-pass filtering, and half-plane phase correlation) improved the signal-to-noise ratio (SNR) from 6.6 to 115.9 on an example image.
  • Higher Correlation Peaks: The phase-correlation peak height was raised from 0.300 to 0.835, substantially higher than previously reported values.
  • Aliasing Causes False Peaks: Rotationally dependent aliasing was shown to create false correlation peaks, sometimes taller than the correct peak, especially at multiples of 90 degrees.
  • Windowing Reduces Degradation: Windowing reduced the degradation caused by aliasing and interpolation errors, leading to significantly better correlation results. For instance, with a low-pass cutoff of N/4, windowed data yielded peak heights of 0.935 and 0.835, compared to 0.667 and 0.452 for unwindowed data.

Limitations and Future Work

  • Trade-off in Filtering: The paper notes a tension between alias reduction (favoring high-pass filtering) and precision preservation (which can be degraded by softening sharp edges with high-pass filtering). This suggests that finding the optimal filtering strategy is a balance.
  • Generalizability: While the experiments are thorough, they are conducted on a single type of image (agricultural aerial photo). Further work could explore the performance on different image types with varying characteristics.
  • Computational Cost: While phase correlation is generally efficient, the paper mentions that some techniques, like Fourier magnitude correlations with many rotations, can be computationally expensive. Future work could focus on optimizing these computational aspects.
  • Further Noise Sources: The paper acknowledges that many other noise sources can impact registration precision in real-world scenarios, and suggests that more experiments are needed to accurately judge the relative qualities of different methods in such complex environments.

Summary updated

BibTeX
@article{stone2003analysis,
  author = {Harold S. Stone and Bo Tao and Morgan McGuire},
  title = {{Analysis of image registration noise due to rotationally dependent aliasing}},
  journal = {Journal of Visual Communication and Image Representation},
  year = {2003},
  doi = {10.1016/s1047-3203(03)00002-6},
}

Loading related content...

Figure