Abstract: In analogy with steerable wavelets, we present a general construction of adaptable tight wavelet frames, with an emphasis on scaling operations. In particular, the derived wavelets can be "dilated" by a procedure comparable to the operation of steering steerable wavelets. The fundamental aspects of the construction are the same: an admissible collection of Fourier multipliers is used to extend a tight wavelet frame, and the "scale" of the wavelets is adapted by scaling the multipliers. As an application, the proposed wavelets can be used to improve the frequency localization. Importantly, the localized frequency bands specified by this construction can be scaled efficiently using matrix multiplication.
Introduction
In signal and image processing, representing complex objects as combinations of simpler "building blocks" like wavelets is a standard practice. However, real-world features—such as cells in microscopy or stars in astronomy—rarely align perfectly with the fixed, dyadic scales (powers of two) used in standard wavelet transforms. While "steerable" wavelets have long addressed arbitrary rotations, handling arbitrary scales continuously is significantly more challenging. Conventional methods often rely on "brute force" multiscale filtering, which is computationally expensive and restricted to discrete scales. This research introduces a framework for "scale-steerable" tight wavelet frames, enabling wavelets to be continuously resized to match local image features through efficient matrix operations.
Methodology
The proposed approach builds upon the foundation of Tight Wavelet Frames. Unlike standard bases, frames are redundant, providing the mathematical flexibility necessary for local adaptation.
Fourier Multipliers and Transformable Families
The construction begins with a "primal" tight wavelet frame, which is extended by applying a collection of Fourier multipliers. If these multipliers satisfy a "partition of unity" (their squared sums equal one), the resulting system remains a tight frame. The core innovation lies in selecting multipliers that form a "transformable family." By utilizing trigonometric polynomials in the log-frequency domain, the authors ensure that a dilated version of a multiplier can be expressed as a linear combination of the original multipliers.
Matrix-Based Scaling and Pseudo-Dilation
Due to this transformability, "steering" the scale does not require new convolutions. Instead, wavelet coefficients are computed once for a set of basis multipliers, and a small steering matrix is then used to find the response at any intermediate scale. The authors also introduce "pseudo-dilation," where the multipliers are scaled while the underlying "mother wavelet" window remains relatively stable, allowing for fine-tuned frequency localization.
Experimental Setup
The method was evaluated using both synthetic and biological datasets:
- Synthetic Data: 1,000 × 1,000 pixel images containing 200 disks with radii varying between 8 and 40 pixels, corrupted with isotropic Brownian motion (fBm) to simulate complex, non-white background noise (autofluorescence).
- Biological Data: Real fluorescence microscopy images of human HT29 colon-cancer cells from the Broad Bioimage Benchmark Collection (BBBC008).
Key Insights and Results
A primary insight is the "Log-Frequency Trick": mapping scale to a logarithmic scale makes dilation equivalent to translation. This allows the use of shiftable trigonometric functions to handle scaling. The redundancy of the tight frame ensures smooth transitions between scales without information loss.
Performance Metrics
The method was compared against Multiscale Laplacian of Gaussian (LoG), the circular Hough transform, and simple thresholding. Key results include:
- Accuracy: The proposed method achieved significantly lower Root Mean Square Error (RMSE) in estimating both the position and radius of spots compared to the Hough transform and LoG.
- Robustness: In high-noise scenarios, the Jaccard Index (measuring detection overlap) remained higher and more stable than competitors, indicating superior noise immunity.
- Efficiency: The approach reaches lower error rates faster than competing algorithms, providing more precise size estimation for a fixed computational budget.
Conclusion
While the method requires higher redundancy—leading to increased memory usage and initial computational costs—it offers unprecedented local scale adaptation. Currently focused on radial multipliers (ideal for circular spots), future work aims to combine scale-steering with traditional rotation-steering to create a fully "affine-steerable" wavelet transform capable of handling rotation, translation, and scaling simultaneously.
Summary updated