The image analysis pipeline

When it applies

First, on the paper and every supplement that will be compared.

How it works

Pulls every embedded image stream out of the PDF, recording for each panel its source document, page, placement rectangle, content hash and a perceptual hash. A pdfplumber fallback covers documents the primary extractor cannot open; it keys panels by stream hash rather than by PDF object number. Supplement containers (.docx and .epub zips, loose image files) are harvested into the same pool.

Inputs

pdf_path

PDF file

max_panelsoptional

integerA per-document budget. Truncation is announced per document, because a truncated supplement and a truncated main PDF bound different claims.

Output

The panel pool, plus counts of what was harvested, skipped and truncated. The supplement scan reads an allowlist of subdirectories rather than walking the tree: the harvest writes its own PNGs into the paper directory, and a general recursion would re-ingest them and manufacture a byte-identical match on every paper in the corpus.

When it applies

Opt-in. A six-panel composite figure is a single embedded stream, so duplication between its own sub-panels is invisible unless the composite is split.

How it works

Finds the gutters — whole rows and columns of near-white pixels — and splits on them, but only when the bands form a REGULAR grid. Regularity, not the mere presence of white lines, is what licenses the split. Columns are measured inside the row bands, because a full-width coloured banner otherwise stops any column reaching the purity threshold. A fallback path exists for montages whose border ring is coloured, which defeats the primary background estimate.

Inputs

segment_montagesoptional

true / falseDefault off.

Output

Sub-panel crops added to the pool, each recorded as a crop of its parent so later stages can tell a crop from an independently embedded image. Measured: turning this on raises the corpus fire rate from 5.0% to 15.1% of papers, mostly through sub-panels that share a plot template rather than content — which is why it is opt-in and why the clique demotion below exists.

When it applies

Across a whole corpus, before the per-paper tiers run. Necessarily two-pass — the set is only knowable corpus-wide.

How it works

An image stream that appears in three or more DISTINCT papers is a journal template, society mark or advert, not a result. No within-paper rule can see this: five Bentham papers in the reference corpus each place one full-page advert twice, via two separate PDF objects — which is exactly the shape of a paper reusing an image as two results.

Inputs

corpus panel index

panel records across many papers

minimum papersoptional

integerThree. Below that, a repeat is not yet evidence of a template.

Output

The excluded hash set plus a report of exactly what was excluded and why, so the exclusion is auditable rather than a quiet absence. Panels carrying an excluded hash are withheld from every tier with a named reason.

When it applies

Always. The strongest tier, because the assertion of sameness is the PDF's own.

How it works

Finds streams the document itself places more than once. This is not a similarity judgement: the file says these two pictures are one object. Either the PDF object number or the stream hash supplies the identity — requiring the object number made this tier silent on exactly the documents the fallback harvest exists to cover.

Inputs

panel pool

harvested panels

Output

Groups of repeated placements. Placements inside captioned figure regions are evidence; reuse living entirely outside figure regions is separated out and capped at indeterminate — page decorations that survived the corpus screen (which needs three or more papers) land there rather than in a finding. Four or more placements of one stream is itself treated as a furniture signature.

When it applies

Distinct panels whose stored bytes hash identically — the same picture embedded twice as two separate objects.

How it works

Compares content hashes of the ORIGINAL embedded bytes. The provenance gate acts here and is the point of the tier: a panel whose pixels we re-rendered ourselves never enters, however its hash falls, and the exclusion is recorded by name. Byte identity is a statement about the authors' files; a re-render is a statement about our renderer.

Inputs

panel pool

harvested panels

Output

Groups of byte-identical panels, and a named list of panels excluded for carrying re-rendered rather than original pixels.

When it applies

Panels that are visually the same picture but not byte-identical — recompressed, resized, or re-exported.

How it works

Computes a perceptual hash per panel and takes connected components of panels within a Hamming distance of 4. Pairs already sharing a non-empty content hash are skipped, because byte identity owns them. That emptiness test is load-bearing: a montage sub-panel carries an empty content hash by contract, so a naive equality check read every pair of crops as “the same bytes” and skipped the entire tier for them.

Inputs

eligible panels

panels passing the texture and glyph gate

Output

Clusters of perceptually near-identical panels. A panel needs only a perceptual hash to enter, not a saved image file — demanding a file made every montage crop inert, excluded from the perceptual tiers and structurally barred from the byte-identity ones, i.e. compared by nothing while the record claimed otherwise.

When it applies

Pairs the first three tiers did not claim — the tier that catches a region rotated, rescaled, mirrored or cropped out of one panel and into another.

How it works

Detects up to 2,000 keypoints per panel, matches descriptors under a 0.6 ratio test, then fits a partial-affine transform with RANSAC at a 3-pixel reprojection tolerance. A match must survive on geometry, not on match count alone: at least 30 inliers, an inlier ratio of 0.25, a recovered scale between 0.2× and 5×, shear under 15°, and keypoints spread over at least 15% of the panel rather than piled in one corner.

Inputs

candidate pairs

panel pairs with loadable pixels

Output

Verified matches with their recovered geometry — scale, rotation, shear, translation and the inlier bounding box — which the classifier and grain stages then interrogate. A panel whose pixels cannot be loaded loses THIS tier only; its perceptual hash still took part in tier 3, and the record says so rather than implying it was never compared.

When it applies

Before the perceptual tiers, on every panel, deciding which are eligible at all.

How it works

Two refusals with different reasons. A panel under 128 pixels on its longest side is treated as a rendered letter or glyph — a panel label, not a result. A panel whose grey-level entropy falls below 3.5 bits, or whose top two intensity bins hold more than 90% of its pixels, has too little texture for perceptual matching to mean anything; western blots live here and wait on a separately calibrated channel. Order matters: the glyph trap is judged first, so a rendered letter is named as what it is rather than as a low-texture blot.

Inputs

panel

a harvested panel

Output

Eligible, or a named exclusion reason carried into the result. This is why a null result must be read alongside the skip counts: panels excluded here were not examined and cleared, they were not examined.

When it applies

On every keypoint match, before it is allowed to count as a duplication.

How it works

Answers two independent questions from the recovered geometry. Is it FURNITURE — a long thin band (a caption strip, running header, axis or scale bar), at aspect ratio 10 or more wherever it sits; or an elongated element at aspect 5 or more that has not MOVED between the two panels, which makes it part of the frame rather than content carried across. Is it PARTIAL — a region duplication covering under half of BOTH panels, in which case the grain test is disqualified. Furniture is never decided by area: the areas of the real and spurious cases overlap and these two signals do not. A 2% area floor is only a backstop for a match too small to be anything at all.

Inputs

match geometry

recovered transform and inlier statistics

Output

A furniture flag and a partial-region flag. Measured: the static-element rule alone reclassified all 45 groups of one paper, a TEM instrument info bar recovering a translation of exactly 0 px at aspect 7.6, where every real duplication in the corpus moved between 278 and 1,057 px. The partial exemption additionally demands 80 inliers, well clear of the 30 a match needs to exist — because the exemption disqualifies grain, and geometry must then carry the finding alone.

When it applies

On whole-panel matches, to separate the same pixels from the same subject photographed twice.

How it works

Aligns one panel onto the other using the recovered transform, subtracts the low frequencies, and correlates what is left — the sensor grain. Genuinely identical pixels carry identical grain; two photographs of the same specimen do not. The measurement is restricted to the bounding box of the matched keypoints rather than the whole warped overlap.

Inputs

aligned panel pair

two panels plus their transform

Output

A correlation coefficient, or null WITH the reason it could not be computed — a discriminator failing silently would leave a group looking merely unclassified, which reads as “we checked and could not say” when nothing was checked. Restricting to the matched region is measured, not cosmetic: on one real case the same pair scores 0.506 over the whole overlap and 0.718 over the matched region, because averaging across surrounding non-duplicated tissue dilutes the statistic. Grain answers “are these the same pixels”; asking it about pixels the match never claimed is the wrong question, which is why the partial-region exemption exists.

When it applies

When montage segmentation is on and sub-panels of one figure match each other.

How it works

If every pair among a figure's sub-panels matches, that is a shared plot template, not duplication. Completeness is the signal: real reuse between two condition panels of one figure is a sparse pair, not a complete graph.

Inputs

match groups and sibling sets

groups plus each figure's sub-panels

Output

Demoted groups, named as template matches. Measured on one paper: 66 groups, exactly the 66 pairs of 12 sub-panels. That is the template. Sparse matches among the same siblings are left standing.

When it applies

On images present in the file that the reader never sees — placed off-page, or never placed at all.

How it works

Separates two facts that wear the same shape in the file. A GHOST is a hidden image that also appears somewhere visible: a float moved pages and the layout engine left a copy behind, so nothing happened. An image present in the file and shown NOWHERE is different — a panel removed late, an earlier version of a figure, something staged and forgotten.

Inputs

hidden and shown panels

panel pool partitioned by visibility

Output

Ghosts recorded as context, so the exclusion is not silent; shown-nowhere images flagged for a human. Neither carries duplication severity, and neither makes a claim about the authors. Reported honestly: measured over 263 corpus papers, 2 papers (0.8%) carry any off-page placement and ZERO carry an image shown nowhere — so the second branch ships as a prospective check with a synthetic control only, rather than implying a yield it has not demonstrated.