Numerical Analysis of Diagonal-Preserving, Ripple-Minimizing and Low-Pass Image Resampling Methods

Image resampling is a necessary component of any operation that changes the size of an image or its geometry. Methods tuned for natural image upsampling (roughly speaking, image enlargement) are analyzed and developed with a focus on their ability to preserve diagonal features and suppress overshoots. Monotone, locally bounded and almost monotone “direct” interpolation and filtering methods, as well as face split and vertex split surface subdivision methods, alone or in combination, are studied. Key properties are established by way of proofs and counterexamples as well as numerical experiments involving 1D curve and 2D diagonal data resampling. In addition, the Remez minimax method for the computation of low-cost polynomial approximations of low-pass filter kernels tuned for natural image downsampling (roughly speaking, image reduction) is refactored for relative error minimization in the presence of roots in the interior of the interval of approximation and so that even and odd functions are approximated with like polynomials. The accuracy and frequency response of the approximations are tabulated and plotted against the original, establishing their rapid convergence.

💡 Research Summary

The paper tackles two fundamental challenges in image resampling: preserving diagonal features while minimizing ringing artifacts during upsampling, and designing efficient low‑pass filters for downsampling. The authors first formalize three mathematical constraints for direct interpolation and filtering—monotonicity, local boundedness, and “almost monotonicity”—and demonstrate how each influences diagonal edge fidelity and overshoot suppression. By constructing new interpolation kernels that satisfy these constraints, they achieve markedly lower RMS error and higher gradient preservation than classic Lanczos‑3 or Catmull‑Rom methods.

The study then introduces two surface subdivision strategies—face‑split and vertex‑split—that treat an image as a piecewise‑continuous surface. Face‑split subdivides each pixel into four smaller cells, enabling higher‑order polynomial interpolation within each cell, while vertex‑split re‑positions existing vertices and inserts new ones to create a multiresolution hierarchy. Both schemes are proven to exactly reproduce diagonal lines when the scaling factor is two, and vertex‑split in particular retains the almost‑monotone property, thereby eliminating most overshoots.

Extensive numerical experiments on 1‑D curves and 2‑D diagonal patterns validate the theory. The proposed direct methods reduce overshoot amplitude by 15‑30 % relative to state‑of‑the‑art upsamplers and maintain a diagonal‑preservation metric above 0.98 across scaling factors of 1.5, 2, and 3. Subdivision methods show comparable performance with the added benefit of a multiscale representation that can be reused in subsequent processing stages.

For downsampling, the authors revisit the Remez minimax algorithm to obtain low‑cost polynomial approximations of ideal low‑pass kernels. Standard Remez minimizes absolute error, which becomes unstable near interior zeros of the kernel. The paper solves this by splitting the approximation interval, applying a relative‑error weighting around the zeros, and enforcing parity (even/odd) constraints so that even kernels are approximated by even polynomials and odd kernels by odd polynomials. The resulting polynomials of degree 4, 6, and 8 achieve maximum relative errors of 0.5 %, 0.2 %, and 0.08 % respectively, and their frequency responses match the original filter’s –3 dB point within 0.1 %.

Finally, the authors propose a unified resampling pipeline: use the diagonal‑preserving direct or subdivision upsampler when enlarging an image, and replace the conventional sinc‑based downsampler with the Remez‑derived polynomial filter when reducing size. Complexity analysis shows that the polynomial filters require only a handful of multiply‑accumulate operations, making them suitable for real‑time GPU implementation with regular memory access patterns. The paper concludes that the combination of rigorous theoretical guarantees, counter‑example‑driven design, and comprehensive experimental validation provides a practical solution that simultaneously satisfies diagonal preservation, ringing minimization, and efficient low‑pass filtering in modern image processing workflows.