Hot Take: Stop Solving PDF Structure at Render Time

TLDR

Inferring PDF document layout by analyzing rendered canvas pixels or raw getTextContent() coordinates introduces non-deterministic behavior dependent on screen DPI, anti-aliasing, and sub-pixel font rendering. Ground-truth PDF structure (tables, borders, zone boundaries) is explicitly encoded in the PDF graphics operator stream (moveTo, lineTo, rectangle, CTM matrices). Processing operator streams deterministically yields scale-invariant layout classification.

Extraction Paradigm	Data Source	Scale Invariance	Deterministic Fidelity
Canvas Pixel Analysis	Rendered Canvas Bitmaps	Fails (Varies by DPI / Zoom)	Low (Anti-aliasing noise)
Operator Stream Parsing	Raw PDF Vector Streams	100% Scale-Invariant	100% Exact Vector Ground Truth

---

Technical comparison: analytical Bezier bounds vs. Subdivision approximation

Using recursive De Casteljau subdivision to approximate Bezier curve bounding boxes introduces arbitrary tolerance cutoffs and creates fragmented line segments:

/ DEFECTIVE: Recursive De Casteljau subdivision generates extra segments and tolerance noise /
function subdivideBezier(p0, p1, p2, p3, tolerance) { ... }
/ RECOMMENDED: Solve Bezier extrema analytically via quadratic derivative roots / export function solveCubicBezierExtrema(a, b, c, d) {   const dA = -3  a + 9  b - 9  c + 3  d;   const dB = 6  a - 12  b + 6 * c;   const dC = -3  a + 3  b;
if (Math.abs(dA) < 1e-6) return [];   const disc = dB  dB - 4  dA * dC;   if (disc < 0) return [];
const sq = Math.sqrt(disc);   const t1 = (-dB + sq) / (2 * dA);   const t2 = (-dB - sq) / (2 * dA);
return [t1, t2].filter(t => t > 0 && t < 1); }

Rule of thumb: Extract document layout directly from graphics operator streams rather than approximating spatial boundaries using rendered canvas pixels.