Appendix A — Glossary¶

This glossary collects the terms the book defines, in alphabetical order. Each entry gives a precise definition in one or two sentences and links to the chapter that develops the notion, where you will find the mechanism, the formula and the measured numbers. Acronyms are listed under their letters, with the expansion in the entry. Where two chapters define the same word differently (for example "calibration" in a recogniser and in a field extractor), the entry says so and links both.

The running example is Folio, the invented claims pipeline of Carrow & Vale Insurance (about 12,000 pages a day, French 60 %, German 20 %, English 15 %, Arabic 5 %). Its targets recur in many entries: CER of at most 2 % on printed text and 8 % on handwriting, amount and IBAN fields at least 99 % correct after validation, and at least 70 % straight-through processing.

Note. An entry is a reminder, not a lesson. If a definition does not give you enough to use the notion, follow the link: every chapter opens with the problem the term solves.

Jump to: A · B · C · D · E · F · G · H · I · J · K · L · M · N · O · P · Q · R · S · T · U · V · W · X

A¶

ABAC (attribute-based access control). An access model that decides each request from attributes of the user, the resource and the context (department, document class, time) rather than from a fixed role. It complements RBAC when "reviewer" is too coarse. Chapter 20.

Abugida. A writing system whose basic unit is a consonant carrying an inherent vowel (Devanagari क is ka), changed by attached marks. It breaks the "one glyph, one character" assumption of recognisers. Chapter 10.

Active learning (pool-based, batch-mode). A labelling strategy in which the learner chooses from a fixed pool of unlabelled examples the \(b\) it would learn most from (the query strategy ranks the pool), labels are bought for those, and the model is retrained once per round until a stopping criterion ends the loop. Chapter 18.

Adapter. The small file of trained low-rank matrices produced by LoRA, a few megabytes, loaded on top of the frozen base model or merged into its weights (\(W \leftarrow W'\)) so that inference costs nothing extra. Chapter 18.

AEAD (authenticated encryption with associated data). One cryptographic primitive (for example AES-GCM) that hides the plaintext and detects any modification of it or of its associated header. It is the default for encrypting stored scans. Chapter 20.

Aliasing. The artefact produced when a signal containing frequencies above the Nyquist limit is sampled anyway: the excess frequencies fold back into the representable range and appear as false patterns (moiré, jagged strokes). Downscaling a page without a low-pass filter aliases the thin strokes of small type. Chapter 1.

Allograph. One of several visual forms of the same letter (a looped and an unlooped cursive "l"). Handwriting models must map many allographs to one class. Chapter 9.

ALTO. Analyzed Layout and Text Object, a Library of Congress XML schema and the standard of digital libraries and newspaper archives. A page is Page > PrintSpace > ComposedBlock > TextBlock > TextLine > String, with explicit SP elements for spaces, a word confidence WC from 0 to 1, and coordinates in pixels, mm10 or inch1200 as declared by MeasurementUnit. Tesseract 5.5.0 writes ALTO v3 down to String only, with no LANG attribute; it is the archive format of the book's pipeline, alongside hOCR and PAGE-XML. Chapter 11.

Amdahl's law. The bound on the speed-up of a job of which only a fraction \(p\) can run in parallel: \(s_n = 1/\big((1-p) + p/n\big)\) with \(n\) workers. It explains why adding workers stops helping once the serial part dominates. Chapter 19.

Angle classifier (textline orientation). A small model that decides whether a cropped line is upright or rotated by 180 degrees, so that the recogniser receives it the right way up. Chapter 8.

Arithmetic intensity. The number of floating-point operations performed per byte moved from memory; it tells whether a model is limited by compute or by memory bandwidth. Chapter 6.

Aspect-preserving resize. Scaling both axes of a line crop by the same factor so letter shapes keep their proportions; the line is then padded to the batch width. Stretching to a fixed width distorts short lines and degrades recognition. Chapter 6.

At-least-once delivery. A queue guarantee that a message is delivered one or more times, never zero, which every practical queue provides and which implies the consumer will sometimes see a message twice. Consumers must therefore be idempotent. Chapter 19.

Autoregressive decoding. Generating a sequence one element at a time, feeding each generated element back as the next input, between a start and an end token. It underlies TrOCR, Donut and every VLM, and brings exposure bias, repetition loops and sequential latency. Chapter 6 and Chapter 7.

B¶

Background (illumination) estimate. An image \(B(u,v)\) holding the paper's brightness everywhere, including under the ink, obtained by removing the thin dark structures with a large morphological closing or a heavy blur. Dividing the page by it flattens shadows and uneven lighting. Chapter 3.

Backpressure. The mechanism by which a full or slow stage makes upstream stages slow down or refuse work instead of accumulating it without bound, typically through a bounded queue that blocks or rejects producers. Chapter 19.

Baseline. The polyline on which the letters of a text line sit, ignoring ascenders and descenders; in handwriting it is curved and varies from line to line. The boundary polygon is the outline of the line's ink region wrapped around it. Chapter 9.

Batch normalisation (BN). A layer that normalises each channel over the batch to zero mean and unit variance and then applies a learned scale \(\gamma\) and shift \(\beta\), which stabilises and speeds up training of deep convolutional recognisers. Chapter 6.

Beam search. A breadth-first decoding that keeps only the \(B\) best partial hypotheses at each step. \(B=1\) is greedy decoding, \(B=\infty\) is exhaustive search; it is approximate because the best full sequence can fall out of the beam. Chapter 2 and Chapter 7.

Bidirectional algorithm (UAX 9). The Unicode algorithm that converts logical order to visual order for a paragraph mixing right-to-left and left-to-right text, by assigning each character a direction class and resolving embedding levels. Needed to compare Arabic OCR output with a reference. Chapter 10.

Binarisation. Turning a gray image into a two-level image: every pixel is declared ink (0) or paper (255). It is a per-pixel classification, done with a global threshold (Otsu) or a local one (Sauvola, adaptive). Chapter 3.

BIO tagging. A token labelling scheme: B-X begins an entity of type X, I-X continues it, O is outside any entity. It lets a model separate two adjacent entities of the same type. Chapter 15.

Bit. The information in one fair yes/no answer. Entropy in bits is the average number of yes/no questions needed to identify an outcome with an optimal strategy; using \(\ln\) instead of \(\log_2\) gives nats (1 nat = 1.4427 bits), the unit of training losses. Chapter 2.

BLEU. A score between 0 and 1 comparing the n-grams of a hypothesis with a reference, with a brevity penalty. Used for formula transcription alongside edit-based metrics. Chapter 13.

Bootstrap. Estimating the sampling variability of a statistic by drawing many new datasets from the observed one with replacement, recomputing the statistic on each and reading the spread. In this book it resamples documents, not characters. Chapter 2, Chapter 8 and Chapter 17.

Born-digital PDF. A PDF produced by software (export, "print to PDF", a report generator) rather than by scanning; its text is usually extractable without OCR, though not always correctly. Chapter 11 and Chapter 12.

Bounding box and polygon. A bounding box is an axis-aligned rectangle \((x_0, y_0, x_1, y_1)\) in pixels; a polygon is a list of four or more points that can follow a rotated or curved line. Chapter 4.

BPE (byte-pair encoding). A way to build a subword vocabulary: start from single characters or bytes and repeatedly merge the most frequent adjacent pair. The vocabulary size sets the sequence length a decoder must generate. Chapter 7.

BufferedImage type. The Java 2D raster layout of an image (TYPE_BYTE_GRAY is one byte per pixel, TYPE_3BYTE_BGR three). Passing the wrong type to Tesseract or to a model is a common cause of garbage output on the JVM. Chapter 21.

Byte-level BPE and byte fallback. In byte-level BPE (GPT-2, RoBERTa, TrOCR-base) the base alphabet is the 256 byte values, so any string can be encoded, but a non-ASCII character becomes two to four bytes that merge into one token only if frequent in training. Byte fallback is the same idea for sentencepiece vocabularies that would otherwise map unknown characters to <unk>. Chapter 7.

C¶

Calibration. A confidence is calibrated when, among all predictions given confidence \(p\), a fraction \(p\) is correct. Tesseract's word confidence, a recogniser's softmax and an extractor's field score are ranking signals until they are calibrated on held-out data; a reliability diagram plots observed accuracy against confidence, and temperature scaling divides the logits by one fitted scalar. See also ECE and refinement. Chapter 2, Chapter 5, Chapter 6, Chapter 15 and Chapter 16.

Candidate set. In post-correction, the short list \(C(x)\) of words a decoder may choose from for an OCR token \(x\). It always contains \(x\) itself, plus every lexicon word within a small edit distance. Chapter 16.

Canonical and compatibility equivalence. Two strings are canonically equivalent when they are the same character spelled differently with no visual or semantic difference (é as one code point or e plus an accent). Compatibility equivalence also merges forms that differ in appearance (ligature ﬁ and fi). Chapter 10.

Catastrophic forgetting. The loss of earlier capabilities when a model is fine-tuned on a narrow set: a model tuned only on garage invoices gets worse on letters. Countered by mixing old data, a small learning rate and adapters. Chapter 18.

Causal mask. The mask in decoder self-attention that forbids a position from attending to later positions, so training with teacher forcing matches left-to-right generation. Chapter 7.

CER (character error rate). The Levenshtein edit distance between the recognised string and the reference, divided by the reference length: \(\text{CER} = (S + D + I)/N\). It can exceed 100 % when the output contains many insertions. Folio's target is at most 2 % on printed text and 8 % on handwriting. Chapter 0, Chapter 1 and Chapter 17.

Character n-gram language model. A model of the probability of each character given the previous \(n-1\), estimated by counting in a corpus. It rescores recogniser hypotheses and scores candidate corrections without needing a word list. Chapter 9, Chapter 10 and Chapter 16.

Claim (lock). A reservation with an expiry that gives one reviewer or worker exclusive handling of a document so that two people do not correct the same page; if the holder disappears, the expiry releases the document for another. Chapter 19 and Chapter 22.

Code point. An integer from 0 to 0x10FFFF that Unicode assigns to one abstract character or mark, written U+0065; each has a name, a general category, a script and a block. An encoding turns code points into bytes. Chapter 10.

Cohen's kappa. Agreement between two raters corrected for the agreement expected by chance: 1 is perfect, 0 is no better than chance. Appendix D.

ColPali and late interaction. ColPali (Faysse et al., 2024) retrieves pages without OCR: a fine-tuned vision-language model emits one 128-dimensional vector per image patch, and the query is embedded as one vector per token. The score is ColBERT's late interaction (MaxSim), \(s(q,d) = \sum_i \max_j \mathbf{q}_i \cdot \mathbf{d}_j\), at the price of about 1,030 vectors per page; it wins on visual pages and loses on exact strings. Chapter 15.

Confidence (field). A score in [0, 1] attached to an extracted field to rank fields from least to most likely correct; it is calibrated only if a score of 0.9 means that 90 % of such fields are right. Chapter 15.

Confidence interval (95 %). A range computed by a procedure that, repeated over many independent samples, would contain the true value in 95 % of them. It is a statement about the procedure, not a 95 % probability for this one interval. Chapter 2 and Chapter 17.

Connected component. A maximal set of ink pixels in which any pixel reaches any other through neighbouring ink pixels. On a binary page, each printed glyph is typically one component, which makes components the building block of classical layout analysis. Chapter 4.

Content-addressed storage. Storing an object under a key derived from its bytes (the SHA-256, for example raw/8f/8f99....jpg), so identical files are stored once and writes are idempotent. A presigned URL is a time-limited URL signed with the server's credentials that lets its holder get or put one named object without seeing the secret key. Chapter 19.

Content stream. The sequence of drawing and text operators of a PDF page. Text is placed by operators, not stored as paragraphs, which is why extraction order may differ from reading order. Chapter 11.

Controller, processor, data subject. GDPR roles: the data subject is the person the data concerns, the controller decides why and how it is processed, the processor processes on the controller's behalf. A hosted OCR vendor is typically a processor. Chapter 20.

Convex quadrilateral. A four-sided polygon whose interior angles are all below 180 degrees, as a photographed rectangle's outline always is. Checking convexity rejects bad page-corner detections before a perspective correction. Chapter 3.

Courtesy amount and legal amount. The courtesy amount is the figure written in digits in the amount box; the legal amount is the same sum written in words. Reading both and checking that they agree (CAR and LAR in cheque reading) is a classic handwriting cross-check. Chapter 9.

CRNN. Convolutional recurrent neural network: a CNN feature extractor feeding a recurrent sequence model, trained with CTC. The output has one element per horizontal position of the line image. Chapter 6.

Cross-attention. An attention layer in the decoder whose queries come from the tokens generated so far and whose keys and values come from the encoder's output. It is how a decoder "looks at" the image. Chapter 7 and Chapter 13.

CTC (connectionist temporal classification). A loss and a decoding rule for sequence labelling when input frames outnumber output symbols and the alignment is unknown. The network emits one symbol or a blank per frame; the loss sums the probability of every path that collapses (merge repeats, then drop blanks) to the transcription. The blank is what allows a genuine double letter. Chapter 0, Chapter 2 and Chapter 6.

Cython binding. A Python extension compiled to C. tesserocr uses one to call Tesseract inside the interpreter, passing images as memory buffers and keeping the model resident, at the price that a crash in the C++ code kills the whole Python process. Chapter 5.

D¶

Data leakage. Information from the test set reaching training or model selection, so the measured error is lower than the error seen in production. Typical routes: the same document in both splits, thresholds tuned on the test set. Chapter 17 and Chapter 18.

DAWG (directed acyclic word graph). A compact automaton that accepts exactly the words of a list; words sharing prefixes and suffixes share nodes. Tesseract keeps three per language (words, punctuation patterns, number patterns) and consults them during the beam search, not after it. Chapter 5.

DBNet++ and adaptive scale fusion (ASF). The journal extension of DBNet that replaces the plain concatenation of the four feature-pyramid levels by a learned weight for each scale and each pixel, so fine scales can dominate on small text and coarse ones on long lines. The paper reports about +4 F-measure on scene-text sets for about 20 % less speed; measure on your own pages first. Chapter 4.

DBNet (differentiable binarisation). A segmentation-based text detector that predicts a probability map \(P\) and, in training, a threshold map \(T\), joined by an approximate step \(1/(1+e^{-k(P-T)})\) so that binarisation is learned. At inference only \(P\) is thresholded, and the contours are dilated ("unclipped") back to the full word size. Chapter 4.

Dead-letter queue (DLQ). Where a message goes after exhausting its retries, so that one poison message cannot block the queue forever and a human can look at it. Chapter 19.

Decoder. The part that turns a recogniser's scores into characters, by CTC collapsing or by attention-based autoregressive generation. Chapter 8, Chapter 6 and Chapter 7.

Decompression bomb. A file that is small on disk and enormous once decoded (a PNG declaring tens of thousands of pixels per side). It exhausts memory in the decoder, so image size limits must be checked from the header before decoding. Chapter 20; see also poison page.

Detector. The stage that outputs boxes or polygons around text without reading them; the recogniser then reads each crop, and the decoder turns the recogniser's scores into characters. Chapter 8.

DETR (DEtection TRansformer). An object detector that predicts a fixed-size set of \(N\) boxes from learned object queries and matches them one-to-one to the ground truth by bipartite (Hungarian) matching, so it needs no anchors and, in principle, no NMS. It is the basis of the Table Transformer. Chapter 12.

Discordant pair. A field or document on which two engines disagree, one right and the other wrong. Only discordant pairs carry information in McNemar's test. Chapter 17.

DLT (direct linear transform). The estimation of a projective map (a homography) from point correspondences by turning the non-linear division into a homogeneous linear system \(A\mathbf{h} = \mathbf{0}\) solved by least squares. Appendix D.

Document-to-Markdown converter. A 2024-2026 tool that takes a PDF and returns Markdown or JSON with headings, tables, formulas and reading order in one call, for LLM and search consumers: Docling (IBM, MIT), Marker, MinerU, PyMuPDF4LLM and PaddleOCR-VL. Licences differ and need reading before shipping (PyMuPDF4LLM is AGPL, MinerU has custom terms). Chapter 8 and Chapter 11.

Domain gap. The difference in error between a model evaluated on held-out synthetic lines and the same model evaluated on real ones. It measures how much a synthetic training set misleads you. Chapter 18.

Donut. Document Understanding Transformer: a Swin encoder and an autoregressive decoder that map a page image directly to a tag sequence convertible to JSON, with no OCR step. See OCR-free model. Chapter 7.

DPIA (data protection impact assessment). The written assessment required by GDPR Article 35 where processing is likely to be high risk (large-scale special-category data, systematic monitoring, automated scoring); Folio qualifies twice, with health data on receipts and fraud scoring. It sits beside the record of processing activities (Article 30), the controller's register of what is processed, why, for whom and for how long. Chapter 20.

DSR (data-subject request). A person's exercise of a GDPR right: access, rectification, erasure, restriction, portability, objection, or not to be subject to a purely automated decision. The controller must answer within one month, so a pipeline must find and delete every copy of a person's document, including derived text and caches. Chapter 20.

E¶

EAST (Efficient and Accurate Scene Text detector). The reference regression-based detector (Zhou et al., 2017): at each cell of a stride-4 map it predicts a score and a box geometry (four edge distances and an angle, the RBOX variant), then merges the many resulting boxes with locality-aware NMS. It outputs rotated rectangles, so it cannot represent curved text, and it regresses very long lines poorly. Chapter 4.

ECE (expected calibration error). Sort predictions into bins by confidence, compare in each bin the mean confidence with the observed accuracy, and average the absolute gaps weighted by bin size: \(\text{ECE} = \sum_b \frac{n_b}{N}\,|\operatorname{acc}(b) - \operatorname{conf}(b)|\). Zero means perfectly calibrated. Chapter 2, Chapter 6 and Chapter 9.

Edit (Levenshtein) distance. The smallest number of single-character insertions, deletions and substitutions that turns one string into another; divided by the reference length it is the CER. Chapter 0 and Appendix D.

Encoding. A rule for turning a sequence of code points into bytes. UTF-8 uses 1 to 4 bytes per code point, with ASCII in one byte. Mixing up bytes and characters is the root of most mojibake. Chapter 10.

Engine diversity. The property that makes combining engines useful: their errors must differ. Sources, from cheapest to most valuable, are another preprocessing of the same image, another model family, another vendor. Chapter 16.

Entity linking (record linkage, entity resolution). Mapping a mention extracted from a document ("Garage Beaumont") to the identifier of an entity in a master table. It has three outcomes, not two: linked (one confident match), ambiguous (several plausible matches) and no match; only the first is automatic. Chapter 15.

Envelope encryption. Encrypting data with a data-encryption key (DEK) and encrypting the DEK with a key-encryption key held in a KMS, so keys can be rotated without re-encrypting the data. Chapter 20.

Epoch, step, batch. A batch is the set of lines processed together; a step is one forward-backward-update on one batch; an epoch is one pass over the training set. Chapter 6.

EXIF. A block of metadata (camera, exposure, date, orientation, GPS position, thumbnail) embedded in JPEG, TIFF and HEIC files. The orientation tag must be applied before OCR, and the GPS tag is personal data to be stripped. Chapter 1 and Chapter 20.

Exposure bias. The mismatch between training with teacher forcing (the decoder always sees the true previous tokens) and inference (it sees its own, possibly wrong, ones). One wrong token can derail the continuation. Chapter 7.

F¶

Fill ratio. For a checkbox, the fraction of pixels inside its inner window that are darker than a threshold. A threshold on it turns a checkbox into a boolean. Chapter 12.

Fine-tuning. Continuing the training of a pre-trained model on a smaller domain set with a small learning rate, so it keeps what it knows and adapts to the new documents. Training from scratch starts from random weights and needs orders of magnitude more data; synthetic pre-training followed by a short real fine-tune is the usual recipe for domain adaptation. Chapter 18.

Formula tree. The structure of a mathematical expression: operator nodes (fraction, sum, script, matrix, radical) with ordered children and symbol leaves. LaTeX or MathML is a serialisation of it, and OCR for maths inverts the rendering. Chapter 13.

Frozen set. An evaluation set whose every file is identified by a hash recorded in a manifest that is itself hashed and committed, so any change, however small, is detected. Chapter 17.

G¶

Gazetteer. A dictionary of known values (first names, streets, garage names) used to find and mask personal data alongside patterns and NER. Chapter 20.

Granite-Docling. IBM Research's 258 M-parameter document model (Apache-2.0): a SigLIP2 encoder, a pixel shuffle that folds 4×4 token blocks into one, and a 165 M Granite language model that emits DocTags (tags for paragraphs, tables, formulas and <loc_> boxes) convertible to Markdown or HTML. A Folio invoice costs 832 image tokens. Chapter 7 and Chapter 14.

Grapheme cluster. What a user perceives as one character, as defined by Unicode Annex 29: a base character plus its combining marks, a Hangul syllable, an emoji sequence. Character counting for CER should count grapheme clusters when scripts require it. Chapter 10.

GriTS (grid table similarity). A table score in [0, 1] from the best order-preserving alignment of two cell matrices, comparing content, topology and location. Chapter 12.

Grounding. Tying generated text to a region of the input image, usually as a bounding box generated by the same decoder as the text. It makes a VLM's output verifiable against pixels. Chapter 14.

Ground-truth (GT) line. A line image paired with its exact transcription, the unit of supervised training for a line recogniser. Model quality is bounded by transcription quality. Chapter 5.

Gutter. The empty vertical band between two columns of text. On Folio's synthetic letters it is 67 px at 200 dpi, much wider than an inter-word gap, which is what column detection exploits. Chapter 4.

H¶

Hallucination. Output text not supported by the pixels. Any decoder with a language prior can write the likely word instead of the printed one; in a VLM it can invent a whole line. A repetition loop is the related failure where the decoder repeats a phrase until it hits the length limit. Chapter 7 and Chapter 14.

Histogram. For an 8-bit image, \(h(k)\) is the number of pixels with value \(k\), \(p(k) = h(k)/N\) the empirical distribution and \(F(k) = \sum_{j \le k} p(j)\) its cumulative form. Otsu's threshold and contrast stretching are computed from it. Chapter 1.

hOCR. XHTML whose class and title attributes carry the OCR structure (ocr_page > ocr_carea > ocr_par > ocr_line > ocrx_word, with bbox 145 137 733 174; x_wconf 91): a browser can display it and a script can parse it. Word confidence x_wconf runs from 0 to 100, and each line carries a baseline and font-size estimates. See also ALTO and PAGE-XML. Chapter 11.

Homogeneous coordinates. Representing a 2-D point \((x,y)\) by \((x,y,1)^\top\), with \((\lambda x,\lambda y,\lambda)^\top\) the same point. Rotation, translation and perspective all become matrix products. Chapter 2.

Homoglyph (confusable). Two different code points with identical or near-identical glyphs (Latin A, Cyrillic А, Greek Α). They defeat string comparison and are a vector for spoofing. Chapter 10 and Chapter 20.

Hough transform (lines). A voting scheme: a line is described by \((\rho, \theta)\), each ink point votes for all lines through it, and peaks in the accumulator are lines. It finds the ruling lines of tables and the skew angle. Chapter 12.

HTR (handwritten text recognition). Recognition of handwriting, usually line by line with a CRNN or a transformer. Chapter 0 and Chapter 9.

I¶

ICR (intelligent character recognition). A name for reading handwriting, used alongside HTR; the book treats both as handwriting recognition, as opposed to OMR, which reads marks. Chapter 0.

Idempotent. An operation is idempotent if doing it twice has the same effect as doing it once. Pipeline stages that may be retried must be, or must carry an idempotency key. Chapter 19.

IDOR (insecure direct object reference). An endpoint such as /task/123 that returns the object to any authenticated user without checking ownership. A frequent flaw in review tools. Chapter 20.

IDP (intelligent document processing). The industry label for the whole ladder from OCR to extraction and decisions. Chapter 0.

IFD (image file directory). In TIFF, a table of tagged entries describing one image (size, compression, pixel interpretation, resolution, location of the pixel data). A multi-page TIFF is a chain of IFDs. Chapter 11.

Image layer. One filesystem diff produced by one build instruction. An image is a stack of layers, shared and cached; ordering instructions from least to most volatile keeps rebuilds fast, and a multi-stage build copies only the finished artefacts into the last stage so compilers and caches never reach production. Chapter 19.

Image token. One position in a language model's input that stands for a block of pixels. It costs what a text token costs, so a high-resolution page can cost thousands of tokens. Chapter 14.

Init-only variable. A Tesseract variable read once when the model loads and ignored if changed afterwards (the dictionary switches, for example). Setting it per call has no effect. Chapter 5.

Ink fraction. The share of pixels inside a box that are darker than a threshold (128 of 255 in the book). Blank paper scores 0, so it is a cheap check that a box a VLM grounded really contains text. Chapter 14.

Interpolation kernel. The weight function used by a resize to compute each output pixel from the input pixels around it (nearest, bilinear, bicubic, Lanczos, area). The kernel decides blur, ringing and aliasing. Chapter 1.

Invisible text layer. OCR text placed over the page image with PDF text render mode 3 (3 Tr): positioned exactly over the printed words and never painted, so the PDF looks unchanged but is searchable. Chapter 11 and Chapter 22.

IoU (intersection over union). For regions \(A\) and \(B\), \(\operatorname{IoU}(A,B) = |A\cap B| / |A\cup B|\), from 0 (disjoint) to 1 (identical). The overlap measure behind NMS and detection scoring. Chapter 4.

J¶

JNA (Java Native Access). A library that lets Java call functions of a shared library without C glue code; Tess4J uses it to reach Tesseract. Chapter 21.

Job. One unit of business work (a document) travelling through a pipeline of stages. Chapter 19.

K¶

KIE (key information extraction). Finding, in a document image or its OCR output, the values of a predefined set of fields (invoice number, total, date, IBAN). It differs from NER in that the layout is part of the evidence, and from full document parsing in that only the schema's fields matter. Chapter 15.

KV cache (key-value cache). The stored keys and values of already-processed positions at every layer, so a decoder computes only the new token's step instead of the whole prefix. It trades memory for speed. Chapter 7 and Chapter 13.

L¶

Lawful basis. One of the six grounds of GDPR Article 6(1) that must be chosen before processing personal data: consent, contract, legal obligation, vital interests, public task, legitimate interests. It determines which data-subject rights apply. Chapter 20.

Layout analysis. Finding the structure of a page: which regions are text, titles, tables, figures, headers, footers, stamps and signatures, and how they nest. Chapter 4.

Layout-aware language model. A transformer encoder pre-trained, with masked language modelling, on documents where each token enters with its text and the box of its word (and in later versions a view of the page image), as in LayoutLM; the objective forces the network to use position as evidence. Chapter 15.

Layout detector. An object detector trained on pages: image in, labelled boxes (paragraph, table, figure, title) out. Chapter 4.

Leakage (evaluation). See data leakage. Chapter 17.

Leptonica. The C imaging library Tesseract uses to read images and to threshold and deskew them. Chapter 21.

Likelihood and prior. In the noisy-channel view, the likelihood \(P(x\mid w)\) is how probable the observed image is if the text were \(w\) (the channel), the prior \(P(w)\) how probable the text is before looking (the language model), and the posterior \(P(w\mid x)\) what the decoder maximises. OCR picks \(\arg\max_w P(x\mid w)P(w)\) (MAP decoding). Chapter 2.

Line-level dataset. Images of single text lines with transcriptions, which removes segmentation so recognisers can be compared on recognition alone. Chapter 9.

Little's law. In a stable system, the average number of items inside \(L\) equals the arrival rate \(\lambda\) times the average time \(W\) an item spends inside: \(L = \lambda W\). It sizes queues and worker counts. Chapter 1 and Chapter 19.

Log domain. Replacing a product of probabilities by a sum of logarithms, so a long sequence's probability does not underflow. Chapter 2.

Logical grid versus visual layout. A table's logical structure is a matrix of rows and columns with spans; its visual layout is how ruling and spacing draw it. They can disagree. Chapter 12.

Logical order and visual order. Logical order is the order text is read, typed and stored; visual order is the left-to-right order glyphs appear on the page. They differ in Arabic and Hebrew. Chapter 10.

LoRA (low-rank adaptation). Fine-tuning that trains small low-rank matrices \(A\), \(B\) added to chosen weight matrices of a frozen model, so \(W' = W + BA\); the result is an adapter. QLoRA does the same on a 4-bit-quantised base model to fit a smaller card. Chapter 18.

LSTM (long short-term memory). A recurrent cell with a cell state updated additively under the control of three gates, which lets gradients flow over many steps; a bidirectional LSTM runs two cells, left to right and right to left. Tesseract 4 and 5 recognise lines with LSTMs. Chapter 5 and Chapter 6.

Luma (and luminance). Luma is a weighted sum of gamma-encoded red, green and blue approximating brightness, \(Y = 0.299R + 0.587G + 0.114B\) (ITU-R BT.601). Luminance is the physical quantity from linear light. OCR works on luma. Chapter 1 and Chapter 3.

M¶

Masked language model (MLM). A transformer trained to recover words hidden from a sentence. Given [MASK] réception de tous, it ranks words for the gap, which scores post-correction candidates in context. Chapter 16.

Matra, virama, conjunct, akshara. In Devanagari, a matra is a dependent vowel sign, the virama cancels the inherent vowel, a conjunct is a consonant cluster drawn as one glyph, and an akshara is the resulting syllable unit. Chapter 10.

Micro versus macro averaging. Micro (pooled) sums all edits and divides by the sum of all reference lengths, so every character weighs the same; macro averages per-document scores so every document does. They can rank engines differently. See also pooled CER. Chapter 8 and Chapter 17.

Minimisation and purpose limitation. GDPR Article 5(1)(b) and (c): collect only data adequate, relevant and limited to a stated purpose. For OCR: do not keep full text or images longer or wider than the task needs. Chapter 20.

Morphological opening. For a binary image \(A\) and a small structuring element \(B\), erosion followed by dilation, \(A\circ B = (A\ominus B)\oplus B\). It removes ink thinner than \(B\); with long horizontal and vertical elements it isolates table rules. Chapter 12.

MSER (maximally stable extremal region). A connected component of the thresholded image whose area barely changes over a range of thresholds. Characters tend to be stable, so MSER is a classical text-candidate detector. Chapter 4.

N¶

NER (named-entity recognition). A statistical model, usually a fine-tuned transformer, that tags spans as person, place, organisation and so on. In privacy work it finds names that no pattern can. Chapter 20.

Noisy channel. A model in which a message \(w\) passes through a medium that corrupts it randomly and the receiver sees only the corrupted \(x\). OCR and post-correction are decoding problems in this model. Chapter 2.

Non-inferiority. Evidence that a new system is worse than the reference by less than a margin \(\Delta\): the one-sided upper bound of the paired difference (new minus old) stays below \(\Delta\). It is the right test when you swap in a cheaper engine; equivalence requires both bounds inside \((-\Delta, \Delta)\). Chapter 17.

Non-maximum suppression (NMS). Sort boxes by score, keep the best, discard every box whose IoU with a kept box exceeds a threshold \(\tau\), repeat on the rest. It turns many overlapping detections into one per object. Chapter 4.

Normalisation (canonical form). A deterministic rewrite that maps equivalent strings to one representative so equal pictures compare equal (LaTeX spacing, Unicode forms). A metric tool, not a correction. Chapter 13.

Normalisation form (NFC, NFD, NFKC, NFKD). A function mapping every string of an equivalence class to one representative. NFD fully decomposes canonically; NFC decomposes then recomposes; the K forms add compatibility mappings. Chapter 10.

Nougat. Neural Optical Understanding for Academic Documents (Meta AI): a Swin encoder and an mBART-style decoder, in the Donut family, that read a whole scientific page and write a markup language with LaTeX for the maths. The small model is CC-BY-4.0, the base model CC-BY-NC-4.0 (non-commercial). Chapter 13 and Chapter 7.

O¶

Object, xref, content stream (PDF). A PDF is a collection of numbered objects (dictionaries, arrays, strings, streams) referring to each other; the cross-reference table (xref) maps numbers to byte offsets. Chapter 11.

OCR engine and recogniser. An engine turns a text image (page or line) into characters; a recogniser handles a single already-detected line; an end-to-end model skips detection and reads the page. Chapter 0 and Chapter 8.

OCR-free model (end-to-end document model). A model that maps a page image directly to structured text (JSON-like tags or markup) with no separate detection or recognition engine, as Donut and Nougat do. It is simple but slow (one sequential token at a time) and gives no word boxes to highlight. Chapter 7.

OCR (optical character recognition). Converting an image that contains text into machine-readable text, returning also where each piece of text is (geometry) and how sure the system is (confidence). Chapter 0.

OEM (OCR engine mode). The Tesseract flag --oem N choosing the recogniser: 0 legacy, 1 LSTM, 2 both, 3 default. It decides how a line is read, not how the page is cut into lines. Chapter 5.

Offered load (erlangs). \(a = \lambda S\), the arrival rate times the service time: the average number of cores kept busy, measured in erlangs, so at least \(\lceil a\rceil\) cores plus headroom are needed. Chapter 22 and Chapter 19.

OmniDocBench. A benchmark of 1,651 PDF pages (10 document types, 5 layouts, 5 languages) that scores text by normalised edit distance, tables by TEDS, formulas by CDM (which compares how a formula renders) and reading order by edit distance. Its overall formula and page count changed during 2025 and 2026, so re-check the version before quoting a score. Chapter 17 and Chapter 8.

OMR (optical mark recognition). Reading checkboxes and bubbles, not characters. Chapter 0 and Chapter 12.

ONNX (Open Neural Network Exchange). A file format for a trained network: a graph of standard operators plus weights, tagged with an operator-set version (opset). ONNX Runtime executes it from Python, Java and C++ without the training framework. Chapter 6, Chapter 19 and Chapter 21.

OpenMP. A shared-memory parallelism standard for C and C++; directives split a loop over threads. Tesseract uses it inside the LSTM, so one process already uses several cores, which matters when you also run several processes. Chapter 5.

P¶

Optical token compression. The idea, behind DeepSeek-OCR, that a page of text is cheaper to hold as pixels than as text tokens if the vision encoder compresses it enough (a 16× convolutional compressor takes a 1,024 px page to 256 vision tokens). The compression ratio is \(r = N_{\text{text}}/N_{\text{vision}}\); the paper reports about 97 % decoding precision for \(r < 10\) and about 60 % at \(r = 20\). Chapter 14 and Chapter 7.

Otsu's method. A global binarisation threshold chosen, from the histogram alone, to minimise the spread of grey levels inside the two resulting classes (ink and paper). Appendix D and Chapter 3.

P¶

PageState. The single typed object that travels through Folio's pipeline: identity (page id, SHA-256), classification results, preprocessing log, recognised lines and decisions. Chapter 22.

PAGE-XML (Page Analysis and Ground Truth Elements). An XML schema from the PRImA group (University of Salford) and the working format of historical-document recognisers (Transkribus, Kraken, eScriptorium): Page > TextRegion > TextLine > Word > Glyph, a polygon (Coords) at every level, an explicit ReadingOrder, typed regions (SeparatorRegion, TableRegion...) and TextEquiv confidences from 0 to 1. The schema allows every level, but Tesseract 5.5.0 writes only regions and lines by default (no Word elements; -c page_xml_level=1 adds them) and never glyphs. Chapter 11.

PARSeq. A scene-text and document recogniser trained with permutation language modelling, which can decode left-to-right, right-to-left or by iterative refinement. Chapter 7.

PARSeq (permuted autoregressive sequence model). A small scene-text recogniser for word crops (ViT-S encoder on a 32×128 image, one-block decoder, about 23.8 M parameters) trained with permutation language modelling. It decodes autoregressively or in one parallel pass, optionally followed by iterative refinement in which each position sees all the others; it reads words, not lines. Chapter 7.

Patch, token. A patch is a \(P \times P\) block of pixels; after a linear projection it becomes a token (a vector) in a vision transformer. Chapter 7.

Pattern plus checksum. A regular expression proposes candidates and an arithmetic check designed into the identifier (modulo 97 for IBANs, Luhn for cards) accepts or rejects them. Precise and language-independent, but limited to identifiers that have a structure. Chapter 20.

PDF/A. The ISO 19005 profile of PDF for long-term archiving: self-contained, fonts embedded, no JavaScript or encryption, no external references. OCR output for the archive is often PDF/A. Chapter 11 and Chapter 22.

Peaky output. A converged CTC network emits each character at one or two frames and the blank everywhere else (---c--o--t---), so frame confidences are close to 0 or 1 and say little about uncertainty. Chapter 6.

Permutation language modelling. Training a sequence model on many generation orders of the same text instead of one, so a single model can decode in several orders. Chapter 7.

Personal data. Any information relating to an identified or identifiable natural person (GDPR Article 4). A scanned claim form is full of it; so are the OCR text, the thumbnails and the logs. Chapter 20.

Pipeline, stage, DAG. A pipeline is a fixed sequence of stages, each consuming the previous one's output; a job is one document travelling through it. A pipeline whose stages branch and merge is a DAG (directed acyclic graph). Chapter 19.

Pixel, sample, channel. A pixel is one position of the grid; a sample is one measured number at one position (a grayscale pixel holds one, an RGB pixel three); a channel is the plane made of one kind of sample over the whole image. Chapter 1.

Point, em. A point is 1/72 inch; the em is the font size in points. In pixels, \(\text{em}_{\text{px}} = \text{pt}\times\text{dpi}/72\), which links font size, resolution and x-height. Chapter 1.

Poison page. An input that makes a worker run far longer than normal or never finish (huge image, page of noise, decompression bomb). A fleet cheap on average is killed by its worst page, hence per-page timeouts. Chapter 5 and Chapter 19.

Pooled CER (ratio of sums). Add all errors, add all reference characters, divide once. Contrast with macro averaging (mean of per-page ratios). Chapter 8 and Chapter 17.

Post-correction. Any processing applied to the text an OCR engine returned, without looking at the image again, aiming to reduce wrong characters or words. Chapter 16.

Power (statistical). The probability that a test rejects the null hypothesis when the engines really differ by a given amount; conventionally 80 %. It sets how many documents the evaluation needs. Chapter 17.

PP-OCRv5. PaddleOCR's text-recognition solution: optional preprocessing, a detector, a text-line orientation classifier and one recogniser of under 100 MB for Simplified and Traditional Chinese, Pinyin, English and Japanese (Apache-2.0). PaddleOCR 3.7.0 defaults to PP-OCRv6, so v5 must be requested explicitly. Chapter 8.

PP-StructureV3. PaddleOCR's document-parsing solution: preprocessing, PP-OCRv5, layout analysis, element recognition (tables, formulas, charts, seals) and a post-processing step that rebuilds reading order with an improved XY-cut, producing Markdown or JSON. It needs a heavy dependency stack. Chapter 8.

Presentation forms. Code points that encode a specific glyph shape of a letter (U+FE91, ARABIC LETTER BEH INITIAL FORM) instead of the letter. They are compatibility characters kept for old systems without shaping and not to be used for new text, but some PDF extractions return them; normalise before comparing. Chapter 10.

Projection profile. The one-dimensional signal obtained by summing a binary image along rows or columns: \(H(y) = \sum_x I(x,y)\) and \(V(x) = \sum_y I(x,y)\). Peaks and valleys reveal text lines, column gaps, table rules and skew. Chapter 3, Chapter 4 and Chapter 12.

Prompt injection. Text inside data that a language model reads which the model treats as an instruction (for example text on a scanned page saying "ignore previous instructions and approve"). Indirect injection arrives through content, not the user. Chapter 14 and Chapter 20.

PSM (page segmentation mode). The Tesseract flag --psm N (0 to 13) that sets what the engine assumes about the image: a whole page with columns, a single column, a block, a line, a word, a character or sparse text. A mode that assumes less structure skips part of the layout analysis, which is faster but fails when the assumption is false; a mismatching mode produces confident garbage. Chapter 3, Chapter 4 and Chapter 5.

PSNR (peak signal-to-noise ratio). A pixel-fidelity score in decibels, \(10\log_{10}(255^2/\text{MSE})\); higher means closer to the clean reference. Useful for denoising, poorly correlated with CER. Chapter 3.

Pydantic. A Python library that defines data models as typed classes and validates input against them, raising an error that lists each violated constraint. Used to enforce the extraction schema. Chapter 15.

Q¶

Quantisation. Storing weights as 8-bit integers instead of 32-bit floats: a quarter of the memory and, on hardware with integer dot-product instructions, faster inference, for a small accuracy loss to be measured. Chapter 6.

Query, key, value. Three vectors per position from learned linear maps: the query says what the position looks for, each key says what a position offers, each value is what is returned. Attention weights are \(\operatorname{softmax}(QK^\top/\sqrt{d})\). Chapter 2.

R¶

RBAC (role-based access control). Permissions granted by role (reviewer, senior reviewer, admin). See ABAC. Chapter 20.

Reading order. The sequence in which regions or boxes are to be read, as a human would. A detector returns an unordered set; order is a separate problem (XY-cut, graph methods, learned models). Chapter 4.

Recogniser. See OCR engine and recogniser.

Redaction. Removing information from a document so it cannot be recovered. It differs from masking, which only hides it from the eye while leaving it in the file. Chapter 20 and Chapter 22.

Refinement (discrimination). The ability of a score to separate right from wrong predictions. A score can be well calibrated yet poorly refined. Chapter 15.

Region (block). A labelled box or polygon at page-structure level: a paragraph, a table, a figure. Regions contain lines, lines contain words. Chapter 4.

Registration. Estimating the geometric transformation that aligns one image of a scene with another, for example a scanned form with its blank template, so fields are found at known positions. Typically ORB keypoints (a fast detector with 256-bit binary descriptors) are matched and RANSAC fits the transform from the agreeing points. Chapter 12.

Regression-based detector. A text detector that predicts at each cell the geometry of its box (edge distances and angle); the decoder is a list of boxes filtered by NMS (EAST is the classic example). Chapter 4.

Review rate. \(1 - s\) where \(s\) is the straight-through rate: the share of documents sent to human review. Chapter 17.

Ringing, blocking, chroma bleeding. JPEG artefacts: ripples around strong edges, visible 8-by-8 block boundaries, and colour smear across edges. They hurt thin strokes at low quality settings. Chapter 1.

ROVER (Recogniser Output Voting Error Reduction). A way to combine several engines (Fiscus, 1997, from speech recognition): align all hypotheses into a word transition network, then vote in each slot with \(S(w) = \alpha N(w)/K + (1-\alpha)\,c_{\max}(w)\), where \(N(w)\) counts the engines that voted \(w\). It beats the best single engine only modestly, and only if the engines' errors differ. Chapter 16.

Run length and RLSA. A run is a maximal sequence of equal pixels along a row. RLSA (run-length smoothing algorithm) fills short gaps of 0s between 1s, merging letters into words and words into lines. Chapter 4.

S¶

Sampling unit. The thing drawn at random in an evaluation. For OCR it is the document, not the page or the character, because pages of one document are correlated; a stratum is a subgroup (document type, source) sampled separately so that rare or costly subgroups get a place in the set. Chapter 17.

Sandbox. An execution environment that restricts what code can do (files, network, system calls). Image and PDF parsers should run in one. Chapter 20.

Schema. The list of fields to extract, each with a name, a type and constraints (total: non-negative amount with two decimals; iban: string passing ISO 13616). Chapter 15.

Scoring profile. A named normalisation applied to both reference and hypothesis before counting edits, stating what is not an error (case, diacritics, presentation forms). Strict, lenient and similar profiles give different CER on the same output. Chapter 10.

Segmentation-based detector. A text detector that predicts which cells are text and recovers instances with connected components or region growing (CRAFT, DBNet, PSENet). Instances are shapes in the map, so curved text is handled better than by a regression-based detector. Chapter 4.

Self-attention. Attention in which queries, keys and values all come from the same sequence. Chapter 7.

Sentencepiece. A tokenisation library that works on raw text, treats the space as a symbol (▁) so it needs no pre-tokenisation, and can use BPE or a unigram model; TrOCR-small's 64,044-entry vocabulary is one. Chapter 7.

Service level indicator and objective (SLI, SLO). An SLI is a measured number describing service quality (share of interactive jobs finished in under 10 s); an SLO is the target for it (99 % over 30 days). The golden signals to watch on any service are latency, traffic, errors and saturation. Chapter 19.

Service time. \(S\), the seconds of one core a page occupies in a worker, from taking the job to writing the decision. Chapter 22.

Shaping. The step, performed by a text-layout engine such as HarfBuzz, that turns code points into glyphs, choosing contextual forms and ligatures. Arabic and Devanagari need it; synthetic data generation too. Chapter 10.

Signal, score, referral. In fraud detection, a signal is one measurable inconsistency (editing trace, duplicated image), a score combines signals, a referral sends the document to an investigator. Chapter 20.

Singular value decomposition (SVD). The factorisation \(A = U\Sigma V^\top\) of any matrix into two orthogonal matrices and a diagonal one of non-negative singular values; the ratio of the largest to the smallest is the condition number, which bounds how much input error can be amplified. It underlies homography estimation and low-rank adaptation. Chapter 2.

Skeleton (confusables). The key defined by Unicode Technical Standard 39 for comparing look-alike strings: normalise, then map each character to a canonical prototype. Chapter 10.

Skew. The small rotation of text lines relative to the image axes, usually under 10 degrees, from a crooked feed or a tilted camera. The book's convention is that the angle \(\theta\) is positive when the text is rotated counter-clockwise, and deskewing rotates by \(-\theta\); it is estimated with projection profiles or the Hough transform. Chapter 3.

Softmax regression. Multinomial logistic regression: class scores \(z_k = w_k^\top x\) become probabilities \(p_k = e^{z_k}/\sum_j e^{z_j}\), trained by minimising the cross-entropy of the true class. Folio's document-type classifier is one, on a 192-cell ink grid. Chapter 19.

Spatial relation. In formula recognition, the geometric rule linking a child symbol to its parent: right of, raised, lowered, above, below, inside. Chapter 13.

Special-category data. GDPR Article 9 data: health, biometric identification, genetic, religion, and so on. Medical receipts in claims fall here. Chapter 20.

Stage contract. The promise a stage makes to its neighbours: which fields of the shared state it reads and writes, and what it does on failure. Chapter 22.

Storage limitation. GDPR Article 5(1)(e): keep identifiable data no longer than necessary. Drives retention rules for scans, text and caches. Chapter 20.

STP (straight-through processing). The share of documents that go from intake to the system of record with no human intervention; \(1 - \text{STP}\) goes to review. Folio's target is at least 70 %. Chapter 0, Chapter 17 and Chapter 19.

Stroke width transform (SWT). An operator that assigns to each pixel of a candidate stroke the width of the stroke it belongs to; text strokes have nearly constant width. Chapter 4.

T¶

Table detection, table structure recognition (TSR), table content recognition. Three steps: locate the table rectangle, recover its logical grid (rows, columns, spans), read each cell. Chapter 12.

Table Transformer (TATR). A DETR-based model, trained on PubTables-1M, that detects tables and recovers their rows, columns and spans; the book runs it on CPU at 0.6 to 1.3 s per table. Chapter 12.

Task prompt token. A special first token of the decoder input telling a multi-task decoder which task to perform (parse, classify, answer). Chapter 7.

Tategaki. Vertical writing in Chinese, Japanese and Korean: characters stacked top to bottom, columns running right to left. Chapter 10.

Teacher forcing. Feeding the true previous tokens to the decoder during training instead of its own predictions, which parallelises training and causes exposure bias. Chapter 7.

TEDS (tree-edit-distance-based similarity). A table score in [0, 1] from the edit distance between the HTML trees of prediction and truth, \(1 - \text{TED}/\max(|T_p|, |T_g|)\). Chapter 12.

Text detection. Finding where text is: boxes or polygons around words or lines, with a confidence. It reads nothing. Chapter 4.

Text line (line strip). A crop containing one line of text with a roughly constant baseline, the input of a line recogniser. Chapter 6.

Threat model. A structured answer to four questions: what we protect (assets), who might attack (actors), where trust changes (boundaries), and what controls answer each credible attack. STRIDE is a checklist of six threat types: spoofing, tampering, repudiation, information disclosure, denial of service, elevation of privilege. Chapter 20.

Token (formula). The unit a model or metric counts; a LaTeX control sequence such as \frac is one token. Chapter 13.

Toolkit. A packaged OCR system you install and call: pipeline code, trained weights, runtime framework and system programs it shells out to. An engine is only its recognising core, and a wrapper (OCRmyPDF around Tesseract) adds a workflow without recognising anything itself. Chapter 8.

ToUnicode CMap. An optional stream attached to a PDF font mapping each code to the Unicode text it stands for. It plays no part in drawing; text extraction depends on it. Chapter 11.

Transfer (of personal data). Moving personal data to a recipient outside the European Economic Area (GDPR Chapter V), which needs an adequacy decision, safeguards such as standard contractual clauses with a transfer impact assessment, or a narrow derogation. Adequacy decisions can be challenged, so check their status; sending scans to a US-hosted API is a transfer. Chapter 20.

TrOCR. A line-level transformer recogniser: an image-transformer encoder (ViT, initialised from BEiT or DeiT) reads a 384×384 line image, and a text-transformer decoder with cross-attention writes subword tokens autoregressively. The base printed model has about 333 M parameters. Chapter 7.

U¶

Under-resourced script. A script or language for which labelled data, fonts, corpora or pretrained models are scarce relative to Latin. Chapter 10.

Undetected-error rate. The share of auto-accepted documents that carry at least one wrong critical field (amount, IBAN, policy number, date of loss). It is the price of straight-through processing. Chapter 17.

Unicharset. The table of all characters a model can output, with properties for each (script, letter, digit, punctuation). A character absent from it can never be produced, whatever the image shows, so adding characters means retraining; in Tesseract a recoder maps each entry to a short code sequence so the last layer need not have one unit per entry. Chapter 5 and Chapter 10.

Union area. The area covered by overlapping boxes, not the sum of their areas; word boxes that share a margin must not count twice. Used for coverage ratios. Chapter 11.

V¶

Utilisation. \(\rho = \lambda S/c\), the fraction of time the servers of a pool of \(c\) are busy. Chapter 19 and Appendix D.

V¶

VGSL (variable-size graph specification language). A compact string, designed for Tesseract's trainer, describing a network layer by layer. Chapter 5.

View (array). An array object that shares the memory of another and differs only in shape, strides and offset. Slicing makes views, so writing through one changes the original. Chapter 1.

Virtual thread. A thread managed by the JVM rather than the operating system (JEP 444, final in Java 21), cheap to create and suited to blocking I/O such as waiting for an OCR service. Chapter 21.

VLM (vision-language model). A network that takes an image and a text instruction and generates text: a vision encoder, a projector and a language model. Strong on layout and semantics, prone to hallucination. Chapter 14.

W¶

WER (word error rate). The same computation as CER over words. Chapter 0 and Chapter 17.

Wilson interval and McNemar test. The Wilson interval is a confidence interval for a proportion that stays inside 0 to 1; McNemar's test compares two systems on the same items using only the discordant pairs. Chapter 13 and Chapter 17.

Window attention and patch merging. Window attention restricts self-attention to non-overlapping windows; patch merging halves the grid between stages. Together they keep a Swin-type encoder affordable on large images. Chapter 7.

Word confidence. Tesseract's 0 to 100 score per word. It is not a probability: a word reported at 90 is wrong far more often than one time in ten. Calibrate it before thresholding. Chapter 5 and Chapter 22.

Word transition network (WTN). A sequence of slots, each holding one vote per engine (a token, or the null arc @ when an engine has nothing aligned there), built by aligning several engines' outputs. Voting picks one arc per slot, and a slot won by the null arc contributes nothing to the output (ROVER). Chapter 16.

Writer-independent split. A train/test split in which no writer appears on both sides, as in the IAM RWTH split. A random split of lines is not: the model meets the test writer's slant in training and scores better than it will on a new customer. Chapter 9.

X¶

x-height, cap height, ascender. The x-height is the height of lowercase letters without ascenders or descenders (x, a, e); the cap height is that of capitals; ascenders rise above the x-height. Below about 12 px of x-height (about 150 dpi for 10-pt text) accuracy drops quickly; 20 to 30 px (about 300 dpi) is comfortable. Chapter 1 and Chapter 3.

XY-cut. A top-down recursive page segmentation (Nagy and Seth, 1984): project ink onto an axis, cut at the widest empty gap, recurse on the pieces alternating axes. Gives reading order for Manhattan layouts. Chapter 4.

See also: Appendix B for commands and parameters, Appendix C for symptoms and fixes, and Appendix D for the derivations behind CER, CTC, ECE and the statistics.