Efficient Rotation-Scaling-Translation Parameters Estimation in Multimodal Settings Based on Fractal Image Model

Mikhail L. Uss, Benoît Vozel, Vladimir V. Lukin, Kacem ChehdiarXiv 2015 · arXiv · OpenReview

Abstract

This paper presents a novel approach to area-based subpixel image registration under the rotation-isometric scaling-translation (RST) transformation hypothesis. Traditional methods often rely on universal similarity measures like mutual information or normalized correlation, which may fail to account for specific textural characteristics. We propose the $ML_{fBm}$ estimator, a parametric model based on fractional Brownian motion (fBm) that adapts to the local texture of image fragments. By employing maximum likelihood estimation, the $ML_{fBm}$ method significantly improves registration accuracy in multimodal settings, even when correlation between images is weak. Experimental results on simulated and real-world remote sensing data demonstrate that the proposed estimator reduces estimation errors by a factor of 1.75 to 2 and decreases the probability of false matches by up to five times compared to state-of-the-art methods.

Introduction

Image registration—the process of aligning two or more images of the same scene—is a fundamental task in remote sensing, medical imaging, and computer vision. While simple translation is often manageable, real-world applications frequently involve complex Rotation-Scaling-Translation (RST) transformations.

Most existing "area-based" registration techniques utilize "one-size-fits-all" similarity measures such as Normalized Cross-Correlation (NCC) or Mutual Information (MI). These methods often struggle with multimodal registration—for instance, aligning hyperspectral data with panchromatic imagery—where sensor noise and varying spectral responses result in weak correlation. This paper addresses these limitations by introducing a parametric approach that models the statistical "roughness" of image textures to achieve superior subpixel alignment.

Methodology: The $ML_{fBm}$ Estimator

The core innovation of this research is the shift from universal measures to a model-based approach using fractional Brownian motion (fBm).

Fractal Modeling

Rather than treating pixels as independent values, the $ML_{fBm}$ estimator models local image fragments as fBm processes. This allows the algorithm to characterize complex textures using two primary parameters: the Hurst exponent ($H$), representing surface roughness, and the texture amplitude ($\sigma_x$). By understanding the statistical structure of the scene, the estimator can more accurately distinguish between true alignment and noise.

Joint Maximum Likelihood Estimation

The algorithm performs a simultaneous estimation of eight parameters:

  • Four RST parameters: Translation ($\Delta t$), scaling ($\Delta s$), rotation angle ($\alpha$), and a secondary translation component ($\Delta r$).
  • Four texture/correlation parameters: Texture amplitudes ($\sigma_{x,RI}, \sigma_{x,TI}$), the Hurst exponent ($H$), and the correlation coefficient ($k_{RT}$).

Key Technical Advantages

  1. Signal-Dependent Noise: Unlike models assuming simple white noise, this method incorporates a signal-dependent noise model, reflecting the realities of modern multispectral and hyperspectral sensors.
  2. Direct Subpixel Estimation: Because the $ML_{fBm}$ log-likelihood function is continuous, the system identifies subpixel peaks directly without the need for pixel interpolation, which often introduces bias.
  3. Global Convergence: To prevent the optimization from settling in local optima, the authors implement a multi-start technique utilizing nine distinct initial guesses.

Theoretical Performance and Confidence

A significant contribution of this work is the derivation of a new Cramér–Rao Lower Bound (CRLB) for RST parameters. This bound provides a theoretical limit on estimation error based on texture roughness, noise levels, and image correlation. In practice, this allows the system to generate a "confidence ellipsoid" for every match, enabling the automatic flagging and removal of unreliable control points.

Experimental Evaluation

The $ML_{fBm}$ estimator was tested against two distinct datasets:

  1. Simulated fBm Data: Ten test scenarios were generated to evaluate performance across varying levels of texture roughness, noise, and correlation.
  2. Real-Life Remote Sensing: The study utilized data from the Hyperion hyperspectral sensor (Band #25) and Landsat 8 multispectral data (Bands B1 and B8). These images were captured 12 years apart with differing spatial resolutions (30m vs. 15m), representing a significant challenge for traditional registration algorithms.

Results and Discussion

The experimental results confirm the superiority of the $ML_{fBm}$ approach:

  • Accuracy: The method reduced estimation errors for translation, rotation, and scaling by a factor of 1.75 to 2 compared to Normalized Gradient Fields (NGF) and Mutual Information (MI).
  • Robustness: The probability of a false match was reduced by up to 5 times. In the Landsat-to-Hyperion tests, the outlier rate for the best-matching fragments was only 10%, whereas traditional methods reached outlier rates as high as 48%.
  • Weak Correlation: The estimator maintained high precision even when the correlation between images dropped to 0.4, making it highly effective for multimodal registration.

Computational Trade-offs

The primary drawback of the $ML_{fBm}$ estimator is its computational intensity. In MATLAB environments, the process is approximately 35 times slower than NCC, taking roughly 20 seconds per fragment pair. This makes the current implementation better suited for high-precision offline applications rather than real-time processing.

Conclusion

The $ML_{fBm}$ estimator represents a significant advancement in subpixel image registration by leveraging the statistical properties of image textures. By moving beyond universal similarity measures to a parametric fractal model, the method provides a robust solution for the challenging task of multimodal RST registration. Future work will focus on reducing computational complexity through GPU acceleration and extending the model to handle anisotropic textures and non-Gaussian noise distributions, such as those found in Synthetic Aperture Radar (SAR) imagery.

Summary updated

Abstract

This paper deals with area-based subpixel image registration under rotation-isometric scaling-translation transformation hypothesis. Our approach is based on a parametrical modeling of geometrically transformed textural image fragments and maximum likelihood estimation of transformation vector between them. Due to the parametrical approach based on the fractional Brownian motion modeling of the local fragments texture, the proposed estimator MLfBm (ML stands for "Maximum Likelihood" and fBm for "Fractal Brownian motion") has the ability to better adapt to real image texture content compared to other methods relying on universal similarity measures like mutual information or normalized correlation. The main benefits are observed when assumptions underlying the fBm model are fully satisfied, e.g. for isotropic normally distributed textures with stationary increments. Experiments on both simulated and real images and for high and weak correlation between registered images show that the MLfBm estimator offers significant improvement compared to other state-of-the-art methods. It reduces translation vector, rotation angle and scaling factor estimation errors by a factor of about 1.75...2 and it decreases probability of false match by up to 5 times. Besides, an accurate confidence interval for MLfBm estimates can be obtained from the Cramer-Rao lower bound on rotation-scaling-translation parameters estimation error. This bound depends on texture roughness, noise level in reference and template images, correlation between these images and geometrical transformation parameters.
BibTeX
@misc{uss2015efficient,
  author = {Mikhail L. Uss and Benoît Vozel and Vladimir V. Lukin and Kacem Chehdi},
  title = {{Efficient Rotation-Scaling-Translation Parameters Estimation in Multimodal Settings Based on Fractal Image Model}},
  year = {2015},
  eprint = {1501.02372},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1501.02372},
}
Figure