1. Synthetic ridges. The ridge flow follows the Sherlock–Monro zero–pole model: θ(z) = θ0 + ½[Σ arg(z−ci) − Σ arg(z−dj)], cores c (Poincaré index +½) and deltas d (−½): loop = 1 core + 1 delta, whorl = 2+2, arch = a close core–delta pair. Ridges grow by iterating oriented Gabor filters on noise (SFinGe style), g = exp(−(s²/2σs² + u²/2σu²)) cos(2πu/P), with a smoothly varying orientation and period P (about 5–6 px) which is what creates endings and bifurcations.
2. Enhancement. Gradients (Sobel) → structure tensor J = Gσ*[gx² gxgy; gxgy gy²], ridge direction θ = ½atan2(2Jxy, Jxx−Jyy) + π/2, coherence (λ1−λ2)/(λ1+λ2) (foreground mask). The period P is estimated from zero crossings of profiles taken across the ridges; the Gabor filter tuned to (θ, 1/P) then enhances and the result is binarised at 0.
3. Thinning and minutiae. Zhang–Suen thinning gives a 1 px skeleton. The crossing number CN = ½ Σi=1..8 |pi − pi+1| over the 8 neighbours is 1 at a ridge ending and 3 at a bifurcation. Spurious points are removed: minutiae within 6 px of the foreground border, short spurs (≤ 8 px), tiny fragments, point clusters (< 5 px) and broken-ridge pairs. The minutia angle follows the ridge (ending) or the stem of the fork (bifurcation, chosen by the orientation field).
4. Matching. Every same-type pair (latent i, template j) votes for a rigid pose Δθ = θj − θi, t = pj − R(Δθ) pi in a Hough accumulator (15° x 8 px bins). The top poses are refined by least squares on the paired minutiae; pairs are formed greedily one-to-one within 8 px and 30°. Score s = n² / (nlatent · ntemplate in overlap).
5. Error rates. Over many genuine (same finger) and impostor (different finger) comparisons: FAR(t) = #{impostor ≥ t}/Ni, FRR(t) = #{genuine < t}/Ng, TAR = 1 − FRR, EER where FAR = FRR, AUC = P(genuine > impostor) + ½P(tie), d′ = (μg − μi)/√((σg²+σi²)/2).
Honest simplifications. Prints are 144x144 px, a small gallery, pure 2D rigid+smooth elastic degradation; real sensors add pores, scars, dryness and nonlinear distortion, real matchers use ridge counts, local minutia structures and learned quality, and real databases are millions of times larger, so a real system’s error rates cannot be read off these numbers. Synthetic prints are far more regular than real ones. Nothing here identifies a real person.