G2-imgImage forensicsChart AnalysisLayer 1 (Deterministic)

Typographic Coherence

Checks whether the text in a chart (labels, numbers, titles) is visually consistent and properly aligned, detecting signs of text that was added or modified after the chart was created.

Technical description

Runs OCR over the chart to obtain text regions with per-region confidence (Tesseract 0 to 100, regions below 30 discarded), detects the axis lines and tick positions, and sums five deterministic signals into a 0 to 5 score (capped): font consistency (variance of OCR confidence across labels, threshold 400, up to 2.0); garbled-text legibility (share of regions below confidence 55 exceeding 0.5, the signature of diffusion-rendered labels, up to 1.5); axis perpendicularity (deviation of the inter-axis angle from 90 degrees above 3 degrees, up to 1.5); tick equidistance (coefficient of variation of inter-tick gaps above 0.15 per axis, up to 1.5 combined); and homoglyph contamination (Latin labels carrying Cyrillic look-alike characters, up to 2.0). Requires the image to be at least 32 by 32 pixels.

How it works

Layer 1 (deterministic). Extracts text regions with position and confidence via Tesseract OCR and detects axes and ticks. Computes the variance of OCR confidence across labels (mixed or pasted text); the share of regions below a legibility floor of 55 (uniformly garbled, generator-rendered text); the deviation of the angle between detected axes from 90 degrees; the coefficient of variation of inter-tick spacing per axis; and the count of labels mixing Latin with Cyrillic homoglyphs from an eleven-character look-alike table. Sums the five contributions, capped at 5.0, and reports each anomaly as a finding.

Why this matters

When chart data is fabricated, labels and values are often edited after the chart is generated, which introduces typographic inconsistencies such as different font rendering, misaligned axes, or text that does not match the chart's native font. Generated charts add a second failure mode: diffusion text-to-image models render labels that are jumbled or barely legible because a locality bias breaks glyph consistency, so OCR confidence falls across the figure. Both traces are visible to OCR and pixel geometry even when they look normal to the human eye.

Score thresholds

0-1: Uniform legible labels in one script, perpendicular axes, evenly spaced ticks, consistent with genuine plotting software
2-3: One signal present: mixed or pasted text, broadly low legibility, a sheared axis, or uneven ticks
4-5: Several signals together, or a homoglyph contamination, consistent with an edited, multiply-sourced, or generator-rendered figure

Limitations

Depends on OCR and axis detection: a figure with no readable text is judged on axis geometry alone, and a figure without clear axes or with fewer than three ticks per axis is judged on its text alone. Thresholds are directional; small or stylised fonts can lower OCR confidence, and rotated or broken-axis designs can raise the geometry signals without manipulation. The homoglyph table covers the eleven most common Cyrillic-to-Latin confusables; Greek and other scripts are out of current scope. Body-text homoglyphs live in the text-module E-series, and scale or axis-semantics incoherence lives in the sibling chart indicator.