Exact Boundary-Moment Layer
Implementation & benchmark of arXiv:2607.11158 — An Edge-Based Formulation for the Exact Computation of High-Order Zernike Moments of 2D Shapes and Images
Source paper: An Edge-Based Formulation for the Exact Computation of High-Order Zernike Moments of 2D Shapes and Images arXiv:2607.11158 ⓘ · analyzed Aug 30, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Add a differentiable layer that maps a polygonal contour or predicted segmentation polygon to high-order complex Zernike moments using exact edge integrals instead of pixel-center sums. Feed the resulting moment vector to a classifier or use it as an auxiliary shape-consistency loss, making the representation insensitive to raster resolution and reducing high-order aliasing.
Formulas
Mathematical statement
The paper defines the Zernike basis as $V_n^m(r,\theta)=R_n^{|m|}(r)e^{im\theta}$, where $n\geq 0$, $|m|\leq n$, and $n-|m|$ is even. Its radial polynomial is $R_n^{|m|}(r)=\sum_{k=0}^{(n-|m|)/2}(-1)^kF(n,m,k)r^{n-2k}$ with $F(n,m,k)=\frac{(n-k)!}{k!((n+|m|)/2-k)!((n-|m|)/2-k)!}$. The moment is $Z_n^m=\sqrt{(n+1)/\pi}\iint_D f(r,\theta)(V_n^m)^*(r,\theta)r\,dr\,d\theta$. For a binary polygonal region $D$ and Cartesian coordinates $x=r\cos\theta$, $y=r\sin\theta$, each integrand $q_{nm}(x,y)=(V_n^m)^*(x,y)$ is a polynomial because powers of $r$ multiplied by $e^{-im\theta}$ can be written using $(x+iy)$, $(x-iy)$, and $x^2+y^2$. Green's theorem gives the boundary reduction: if $A_{nm}(x,y)=\int_0^xq_{nm}(s,y)\,ds$, then $\iint_Dq_{nm}(x,y)\,dx\,dy=\oint_{\partial D}A_{nm}(x,y)\,dy$. For a straight edge $(x(t),y(t))=(x_i+t\Delta x_i,y_i+t\Delta y_i)$, the edge contribution is $\int_0^1A_{nm}(x_i+t\Delta x_i,y_i+t\Delta y_i)\Delta y_i\,dt$. The transformed integrand is polynomial in $t$, so sufficiently high-order Clenshaw-Curtis or Gauss-Legendre quadrature is exact up to floating-point error.
Implementation notes
Integrate this at the feature-extraction boundary between a polygonal mask or contour head and the downstream classifier. Inputs are batched vertices $v_i=(x_i,y_i)$ normalized into the unit disk, with counterclockwise ordering; optionally rescale each object by its predicted or ground-truth bounding box. Precompute polynomial coefficients for every requested $(n,m)$ using the factorial formula. Convert each complex basis term into a Cartesian polynomial $q_{nm}(x,y)$ and precompute its antiderivative $A_{nm}(x,y)=\int_0^xq_{nm}(s,y)\,ds$. For every edge, compute $x(t)=x_i+t(x_{i+1}-x_i)$ and $y(t)=y_i+t(y_{i+1}-y_i)$ at fixed Clenshaw-Curtis nodes, evaluate $A_{nm}(x(t),y(t))$, multiply by $\Delta y_i$, and sum with quadrature weights. Multiply by $\sqrt{(n+1)/\pi}$ and concatenate real and imaginary parts. Use automatic differentiation through vertex coordinates. Enforce polygon orientation or multiply by the signed-area orientation. The mathematical quantities computed exactly are the basis coefficients, Cartesian polynomial, antiderivative, and boundary integral; estimate only the moment cutoff $N$, quadrature order $Q$, and normalization strategy. First test on MNIST or synthetic polygons with a small CNN augmented by a 20-dimensional moment vector. Compare against pixel-center moments at image resolutions 16, 32, and 64 under random rotation, rescaling, and rasterization. Measure classification accuracy, reconstruction error, moment variance across resolutions, gradient norms, and runtime. Success means higher cross-resolution accuracy and lower high-order moment variance at equal feature dimension without a substantial training-time increase.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.