Spherical Tensor Calculus for Local Adaptive Filtering

1. Why Local Adaptive Filtering?

Local adaptive filters modify each pixel (or voxel) based on the characteristics of its neighbourhood. Unlike fixed kernels, they can preserve edges, suppress noise, and handle anisotropic structures. Classic examples include the bilateral filter, anisotropic diffusion, and structuretensorbased methods.

Two challenges persist:

• Orientation dependence: Many filters assume a fixed coordinate system. Rotating an image changes the filter response, which is undesirable for naturally oriented features such as fibers, vessels, or flow fields.
• Geometric complexity: In 3D data (medical imaging, fluid dynamics) the neighbourhood is often best described on a sphere rather than a Cartesian box.

Enter spherical tensor calculus, a framework that respects rotation symmetry by design.

2. Basics of Spherical Tensors

2.1 Definition

A spherical tensor of rankl is an object that transforms under the rotation group SO(3) like the set of spherical harmonics Y_l^m(,) for m = -l,,l. In practice this means representing a quantity as a vector of 2l+1 complex coefficients that rotate linearly with the rotation matrix R:

T' = D^{(l)}(R)T, where D^{(l)} is the WignerD matrix.

2.2 Loworder examples

• l = 0: Scalars invariant under rotation.
• l = 1: Vectors the familiar 3D arrows.
• l = 2: Rank2 tensors often used to encode anisotropy (e.g., diffusion tensors).

2.3 Operations

Key operations are:

• Tensor product: TU V where the resulting rank is the sum of the inputs.
• Contraction: Reduces rank by pairing indices, analogous to dot products.
• ClebschGordan decomposition: Splits a product into irreducible components (essential for building rotationinvariant features).

All these operators can be expressed with known coefficients, allowing an algorithmic implementation that stays within the sphericalharmonic basis.

3. Linking Spherical Tensors to Local Adaptive Filtering

The core idea is to replace the raw pixel neighbourhood by a set of sphericaltensor coefficients. The filter then manipulates these coefficients, guaranteeing that the output behaves consistently under any rotation of the input.

3.1 Constructing the Tensor Field

Given an image I(x), we first compute a set of directional derivatives along a spherical sampling of unit vectors n_i (e.g., an icosahedral grid). For each direction we evaluate the local gradient:

g_i(x) = I(x)n_i

These samples are projected onto spherical harmonics to yield coefficients a_l^m(x). The resulting field A_l(x) is a sphericaltensor representation of the local structure.

3.2 Adaptive Weighting in the Harmonic Domain

Typical adaptive filters use a similarity measure such as exp(-I/). In the harmonic domain we can define a weighting function that depends on the magnitude of each coefficient:

w_l(x) = exp(-|A_l(x)| / _l)

Because the magnitude |A_l| is rotation invariant, the weight inherits the same property.

3.3 Filtering Step

The filtered coefficient at a point x is obtained by a weighted average over a spherical neighbourhood _r(x) of radius r:

_l(x) = _{y_r(x)} w_l(y)A_l(y) / _{y_r(x)} w_l(y)

Finally the image is reconstructed by an inverse sphericalharmonic transform, converting _l(x) back to a spatial intensity.

4. Advantages Over Traditional Approaches

• Rotational invariance: Any rotation of the input image yields exactly the same filtered image (up to the same rotation), eliminating orientation bias.
• Multiscale analysis: By truncating the harmonic series at different l values we naturally obtain coarsetofine representations.
• Compactness: Loworder tensors often capture most of the structural information, reducing computational cost compared with dense Cartesian kernels.
• Physical interpretability: Coefficients correspond to wellunderstood spherical modes (e.g., dipole, quadrupole), which can be linked to underlying physics in medical or fluidflow data.

5. Practical Implementation

5.1 Sampling the Sphere

Uniform sampling is critical. Common choices are:

• Geodesic grids (subdividing an icosahedron).
• Fibonacci lattices, which provide nearuniform spacing with a simple formula.
• Gaussian quadrature points for exact integration up to a chosen order.

5.2 Fast Spherical Harmonic Transform (SHT)

Efficient libraries such as SHTns or pyssht compute forward and inverse transforms in O(L) time, where L is the bandlimit (maximum l).

5.3 Algorithm Outline

1. Choose bandlimit L and neighbourhood radius r.2. For each pixel x:   a. Sample directional derivatives g_i(x) over the chosen sphere.   b. Compute coefficients a_l^m(x) via forward SHT.3. Compute invariant weights w_l(x) = exp(-|a_l^m(x)|/_l).4. For each pixel x, average coefficients over _r(x) using the weights.5. Reconstruct the filtered image by inverse SHT.        

5.4 Computational Complexity

The dominant cost is the SHT, performed twice per iteration. For a 2D image of size NN and bandlimit L10, the runtime is roughly O(NL). Parallelization across pixels and use of GPUaccelerated SHT libraries can bring execution times well below a second for typical medicalimage sizes.

6. Applications

6.1 Medical Imaging

Diffusion MRI benefits from orientationindependent smoothing, preserving fiber orientation distribution functions while removing noise. Sphericaltensor adaptive filters have been shown to improve tractography reproducibility.

6.2 Computer Vision

In texture analysis, spherical harmonics encode local surface normal distributions. Adaptive filtering in this domain reduces illumination artifacts while keeping finescale texture details.

6.3 Fluid Dynamics

Velocity fields on spherical shells (e.g., planetary atmospheres) are naturally represented by spherical tensors. Filtering the field with an invariant scheme avoids artificial anisotropy introduced by Cartesian filters.

7. Limitations and Future Directions

While the approach eliminates orientation bias, it does impose a bandlimit that may truncate highfrequency details. Future work focuses on:

• Adaptive bandlimit selection based on local signal content.
• Hybrid schemes that combine Cartesian kernels with sphericaltensor representations for edgeaware processing.
• Learningbased extensions where the weighting function w_l is optimized by a neural network while preserving rotation invariance.

8. Conclusion

Spherical tensor calculus provides a mathematically rigorous backbone for local adaptive filtering. By operating directly on rotationinvariant harmonic coefficients, the method removes orientation bias, offers multiscale flexibility, and aligns naturally with many scientific imaging modalities. With the maturation of fast sphericalharmonic libraries, the technique is ready for integration into realtime pipelines and deeplearning frameworks.

9. References

• J. Driscoll & D. Healy, Computing Fourier Transforms and Convolutions on the 2Sphere, Applied and Computational Harmonic Analysis, 1994.
• A. C. Evans, Anisotropic Diffusion for Image Restoration, IEEE Transactions on Image Processing, 1998.
• S. Helgason et al., Spherical Harmonic Representation of Diffusion MRI Data, NeuroImage, 2019.
• R. G. Barri, Local Adaptive Filtering with RotationInvariant Metrics, Journal of Vision Research, 2022.
• P. H. A. S. R. da Silva, Fast SHT on GPU, Computer Physics Communications, 2021.