Abstract
This paper presents a rigorous mathematical framework for enhancing and completing contours in 2D images by lifting them into the 3D Euclidean motion group $SE(2)$. We derive explicit solutions for linear, left-invariant (convection)-diffusion equations, which serve as forward Kolmogorov equations for stochastic processes. By utilizing invertible orientation scores, we demonstrate that complex image structures, such as crossing fibers in medical imaging, can be effectively denoised and completed. Our results show that geodesics on $SE(2)$ provide a computationally efficient and accurate alternative to traditional elastica for contour completion.
Introduction
A fundamental challenge in image processing is the enhancement of elongated structures, such as blood vessels or collagen fibers, especially in the presence of noise and interruptions. Traditional isotropic filters, like Gaussian blurring, tend to destroy thin lines by smoothing in all directions equally. Furthermore, when structures cross in 2D, local orientation becomes ambiguous, leading to artifacts in standard enhancement algorithms.
This work addresses these issues by modeling image data on the Lie group $SE(2)$, the group of rotations and translations in the plane. By "lifting" a 2D image into a 3D orientation score, we can separate crossing structures in the orientation dimension and apply left-invariant diffusion processes that respect the underlying geometry of the contours.
Methodology
Invertible Orientation Scores
The core of our approach is the transformation of a 2D image $f(x,y)$ into a 3D orientation score $U(x,y,\theta)$. This is achieved through a specialized unitary wavelet transform. Unlike standard transforms, this "lifting" is fully invertible, allowing us to process the image in 3D and then collapse it back to 2D without information loss.
As shown in Figure 1, the orientation score captures local directional information. The kernels derived for these scores (panels c-h) are anisotropic, allowing for smoothing along the contour while preserving its sharpness.
Left-Invariant Evolution Equations
We model contour enhancement and completion as stochastic processes. The evolution of the probability density of a "random walker" in $SE(2)$ is governed by the forward Kolmogorov equations. To ensure that our processing is independent of the image's initial orientation or position, we require these equations to be left-invariant.
Figure 4 illustrates the concept of left-invariance. A tangent vector $X_e$ at the identity element $e$ is mapped to a tangent vector $X_g$ at any group element $g$ via the left-translation $L_g$. This geometric consistency ensures that the diffusion process is intrinsically tied to the group structure of $SE(2)$.
Explicit Solutions and Geodesics
A major contribution of this paper is the derivation of explicit Green's functions (heat kernels) for these evolution equations. These solutions involve Mathieu functions, which describe the spread of "ink" in the position-orientation space.
Furthermore, we compare two methods for contour completion: Mumford’s direction process (minimizing bending energy, or elastica) and geodesics on $SE(2)$. We demonstrate that geodesics provide nearly identical results to elastica but are significantly easier to solve analytically, making them more suitable for practical implementation.
Applications and Results
The proposed framework, termed Coherence Enhancing Diffusion on Orientation Scores (CED-OS), was tested on various datasets:
- Synthetic Data: Validated curvature estimation and kernel accuracy on spirals and concentric circles.
- Medical Imaging: Enhanced bone tissue and muscle cell images. Our Hessian-based approach on orientation scores provided more accurate curvature estimates on noisy data compared to standard structure tensor methods.
- Crossing Fibers: Successfully processed collagen fibers in heart tissue. While standard 2D Coherence Enhancing Diffusion (CED) fails at fiber crossings, our 3D approach maintains the integrity of each fiber by separating them in the orientation dimension.
Conclusion
By integrating Lie group theory with stochastic differential equations, we provide a robust and mathematically sound method for contour processing. The derivation of explicit solutions on $SE(2)$ and the use of invertible orientation scores represent a significant advancement over traditional 2D methods, particularly for complex medical imaging tasks where preserving structural topology is critical. Future work will focus on extending this framework to $SE(3)$ for volumetric data.
Summary updated