Rectifying Higher-Order Convolution Kernels: Discrete Expansion of Directional Derivative Integration and Multi-Scale 8-Ray Star-Perimeter Architectures
Year: 2026
Type: preprint
Abstract: Smoothness is uniquely defined as the integral of the first-order directional derivative over the unit circle. As geometric corners are ground wider with an increasing radius of curvature, turning features do not vanish; instead, they retreat into higher-order Taylor expansion terms, causing traditional low-order and Laplacian kernels to misinterpret presence as absence. This note expands the first-order directional derivative integral into an accumulated series across all orders and establishes a unified layer-by-layer testing criterion: accept present non-zero responses, or treat vanished lower orders as zero and advance to the next order. Furthermore, we compare fixed \(3 \times 3\) windows against expanded \((2m + 1) \times (2m + 1)\) windows (\(n\)-point forward difference formula) and propose a multi-scale 8-ray star-perimeter ("Star + Box") architecture. By combining direct integer grid sampling along the 8 principal rays with piecewise linear interpolation along the perimeter box, this framework achieves pure feature extraction with exactly one value per frame while eliminating angular blind spots. (Note: This work extends the prior paper "Rectifying Smoothness via Directional Derivative Integration", DOI: 10.5281/zenodo.24223693).
Author: Wu, Mengnien
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236568006
