Finds charts in PDF files and recovers the analytical formula of every curve on them.
No neural networks anywhere: the whole pipeline is deterministic, reproducible and explainable — every number in the output can be traced back to a specific geometric feature of the page.
$ analyze_pdf paper.pdf
Page 1 — source: vector PDF graphics
Chart detected, confidence 0.96.
X axis: "X", linear scale, range 0…10, 6 ticks, calibration R² 1.0000
Y axis: "Y", linear scale, range 0…50, 6 ticks, calibration R² 1.0000
Series 1 "linear A" (line, blue, 200 points), X ∈ [0; 10], Y ∈ [0.9868; 20.99]
FORMULA: y = 2·x + 0.9868
model "linear", R² = 1.00000, RMSE = 5.774e-13, 2 params
Series 2 "quad B" (line, red, 200 points), X ∈ [0; 10], Y ∈ [-0.01318; 49.99]
FORMULA: y = 0.5·x^2 + 0.0002635·x - 0.009103
model "parabola", R² = 1.00000, RMSE = 0.002764, 3 params
Series 3 "sine C" (line, green, 200 points), X ∈ [0; 10], Y ∈ [16.99; 32.98]
FORMULA: y = 7.993·sin(0.8·x - 0.0007325) + 24.99
model "sine", R² = 1.00000, RMSE = 0.007426, 4 params
Text extracted from the PDF (axis titles, curve labels) is of course reproduced in whatever language the document uses.
- Decides whether the page contains a chart at all — weighted score over: two long perpendicular lines, short tick strokes touching them, numeric labels along the axes that fall on a straight line under regression, grid lines, and a polyline with many nodes inside the axes box. The decisive feature is the linearity of the labels: for random text the regression R² is low, for a real axis it is ≈ 1.
- Calibrates the axes — pixel → value regression with iterative worst-point rejection.
A logarithmic-scale hypothesis is tested separately (same regression over log10(value)); this matters more than it sounds, because a straight line on a semi-log axis is an exponential, and without detecting the scale the formula comes out meaningless.
- Extracts every curve and converts it to data coordinates.
- Fits a formula — 11 models, winner picked by parsimony/AICc rather than by max R².
Two independent front-ends feed step 3, chosen automatically:
- Vector (src/vector.cpp) — the main path. In a PDF a chart is stored as paths and text, so curve coordinates are read out of the file exactly, with no computer vision. Accuracy: fractions of a percent.
- Raster (src/raster.cpp) — for scans and embedded images. Axes are found by morphological opening with a long kernel, labels are read with Tesseract, the curve is isolated by saturation/hue (for black curves: dark pixels minus long straight lines, i.e. minus grid and frame), then a per-column median gives the trace. Measured accuracy on the test scan: ≈ 0.3 % of the range.
When several curves share a chart, each series gets its own label:
- legend — if a short coloured swatch sits immediately left of a text run, the label is assigned to the series of that colour, not to the geometrically nearest curve (a legend usually sits in a corner, so "nearest curve" would hand every entry to whichever curve happens to pass by it);
- label next to the curve — otherwise the nearest series is taken, within 15 % of the shorter side of the plot box.
Matching is one-to-one and greedy by increasing cost. Text runs already consumed as axis numbers, axis titles or the chart title are excluded from the candidates. Vector branch only — see Limitations.
11 models: polynomials of degree 1–5, exponential, power, logarithm, sine, logistic,
Gaussian, hyperbola, square root. Each gets a meaningful initial guess (log-linearisation
for exponential and power, FFT peak plus mean-level crossing count for the sine, half-maximum
position for the logistic) — with p0 = {1,1,1} almost nothing converges.
The winner is not the maximum R². By R² a high-degree polynomial always wins, because it eats the noise and the discretisation error. The rules, in order:
- if several models reach R² ≥ 0.9999 — the one with fewer parameters wins;
- otherwise, among models whose RSS is no worse than 1.6× the best — again fewest parameters;
- inside that group — by AICc.
On synthetic data (11 dependency types × 2 noise levels) this rule scores 22/22.
Dependencies (Ubuntu 24.04):
apt-get install cmake ninja-build pkg-config \
libmupdf-dev mupdf-tools libeigen3-dev libceres-dev \
libgflags-dev libgoogle-glog-dev \
libfreetype-dev libjpeg-dev libjbig2dec0-dev libopenjp2-7-dev \
libharfbuzz-dev libgumbo-dev libmujs-dev \
libopencv-dev libtesseract-dev tesseract-ocr tesseract-ocr-rustesseract-ocr-rus is only needed to read Cyrillic axis titles; everything else works
without it.
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)Produces build/analyze_pdf.
./build/analyze_pdf chart.pdf # human-readable report
./build/analyze_pdf chart.pdf --json # machine-readable
./build/analyze_pdf scan.pdf --csv out/ # also dump curve points as CSV
./build/analyze_pdf chart.pdf --force raster # force the CV path
./build/analyze_pdf chart.pdf --force vector # force the vector path
./build/analyze_pdf scan.pdf --dpi 300 # render resolution for the raster pathTesseract prints its own diagnostics to stderr; stdout stays clean, so --json can be piped
directly into a parser.
pdf_backend.hpp and raster.hpp are the only places that know about MuPDF and
OpenCV/Tesseract respectively; the rest of the code works with their plain structs
(PageContent, RawPath, TextSpan, PdfDocument::Raster).
reference/ holds the fixture PDFs plus two recorded outputs:
- expected_cpp.txt— what this implementation prints on all nine fixtures. Regenerate and diff it to catch regressions.
- expected.txt— the original Python implementation's output, in Russian. Kept for provenance; useful for comparing numbers, not text.
for f in exp sin logy scatter_parabola power_en no_chart raster_exp multi_text multi_legend; do
echo "########## $f.pdf"; ./build/analyze_pdf reference/$f.pdf 2>/dev/null; echo
done > /tmp/out.txt
diff /tmp/out.txt reference/expected_cpp.txt && echo "no regressions"- Curve labels are vector-only. The raster path does not look for them yet — that needs OCR over the whole plot area rather than the narrow strips next to the axes, and it would keep catching the grid and the curves themselves.
- Same-coloured overlapping curves are not separated — they merge into one series.
- Closed and parametric curves (circle, hysteresis loop) are detected, but a y(x)formula is meaningless for them; the report flags the X-ambiguity.
- Bar and pie charts are recognised as "a chart", but the dependency model does not apply to them.
- Complex functions outside the 11-model library (sums of harmonics, damped oscillation, piecewise definitions) are not recognised as such — the tool still reports the best of the 11, just with a lower R². There is no explicit "I don't know this shape" signal beyond that R².
- Extrapolation past the plotted range is unreliable — the model was only fitted inside the visible window.
- Cyrillic in labels. matplotlib writes PDFs with Type3 fonts that carry no ToUnicode
map, so the text layer returns garbage for Cyrillic. Handled by re-reading the title with
OCR off a page render, which needs tesseract-ocr-rus.
- Lost minus sign. The same Type3 fonts often drop the minus glyph, so an axis
−4 −2 0 2 4extracts as4 2 0 2 4. Handled by testing "first/last k labels are negative" hypotheses and keeping the best R².
Both cost real debugging time and are not obvious from the MuPDF docs.
- Do not flip the page coordinates yourself. fz_bound_page/fz_run_pagealready hand you a page space whose origin is top-left with y growing downwards — unlike the raw coordinates insidefz_path, whichfz_path_walkersees before thectmis applied. The transform you pass should therefore be a pure shift,fz_make_matrix(1,0,0,1,-x0,-y0); adding a flip mirrors the whole page.
- Merge fill_path+stroke_pathfor the same path. The PDF operatorB(fill and stroke) reaches anfz_deviceas two separate callbacks with the samefz_path*and the samectm. PyMuPDF'sget_drawings()reports this as a single object (type: "fs", withfillandcolortogether). Without merging them, the axes frame is counted twice and the grid-line/tick statistics in the detection score come out inflated.
- Solver — Ceres instead of scipy.optimize.curve_fit. The winning model's formula and R² match the reference byte-for-byte almost everywhere; 2nd/3rd place in the "alternatives" list occasionally differs, because on deliberately bad models (a Gaussian fitted over a sine) Ceres converges to a different local optimum than scipy's LM. This never changed the winner in testing.
- --force vectorreally means vector-only. In the original,- forcewas only branched on for- "raster";- "vector"did not disable the automatic raster fallback, which contradicted its own CLI help.
- Per-curve labels — new, the original identified series only by index and colour.
- The report is in English (the original printed Russian). The translation was verified
by hashing every numeric token in the output before and after: identical, so only wording
changed. As a side effect the report can no longer be byte-compared against the Python
reference — hence the separate reference/expected_cpp.txtbaseline.
AGPL-3.0-or-later — see LICENSE.
This is dictated by the dependency on MuPDF, which is AGPL (or a paid commercial licence
from Artifex). Everything else here — Eigen (MPL2), Ceres (BSD), OpenCV (Apache-2.0),
Tesseract (Apache-2.0) — is compatible with a more permissive licence. If you need one,
replace the MuPDF backend with PDFium (BSD): the PDF-specific code is confined to
src/pdf_backend.cpp behind the interface in include/plotparse/pdf_backend.hpp.