ocr_train — SYMBOL `ocr` op

• Data kinds: none → table (an op determined by its arguments alone — it takes no image or data input)

• Call: import fullseye as fs; fs.ledger.ocr_train(font='dot5x7', char_height_px=(16, 20, 26, 34), variants=None, seed=0, neighbour='random') -> 'dict' (to call the implementation directly, import ocrclassic; ocrclassic.ocr_train(font='dot5x7', char_height_px=(16, 20, 26, 34), variants=None, seed=0, neighbour='random') -> 'dict'; from the registry, opssymbol.get("ocr_train"))

• GPU: none (CPU reference implementation only)

Usage

Render the font under nuisance variations and collect templates (features + labels) and the rejection thresholds (table).

> The detailed description below is the original text — the summary and the headings are translated.

Render every character under a small deterministic set of nuisance variations and collect feature templates for k-NN.

各試料は「乱数(seed 固定)で選んだ隣字 + 字 + 隣字」の 3 字行として描き、撹乱(字高・にじみ・ぼけ・小数ずれ)をかけてから、読み取りと同じ規則で中央の字を切る: 字間の中点(真値)に最も近い列射影の極小を切れ目とし(読みの切れ目候補と同じ規則)、その間の字の画素(ink >= 0.5)の外接箱から、重心中心・送り幅の窓で特徴量を取る。隣字を固定(例 "H")にすると、読みの場で隣に来る字のにじみが学習に無く、ぼけ 0.1h で正解率が 2 割落ちた(実測)ので乱数にしている。切れなかった試料(字の画素が無い)は真値の箱で代替し、件数を n_fallback に残す。

Each sample is a 3-glyph line (seeded random neighbours + glyph + neighbour); the middle glyph is cut with the same rule as reading (profile minima nearest to the true gap midpoints) and featurised with the centroid-centred pitch-wide window. A fixed neighbour lost ~20 % accuracy at blur 0.1 h (measured), hence random neighbours.

Parameters

font : 書体名か書体辞書。

char_height_px : 字高の組(各 8..256)。既定 (16, 20, 26, 34)。

variants : {"ink_gain": tuple, "blur": tuple (σ / 字高), "subpixel": bool, "noise": float}。

None で既定(にじみ 3 x ぼけ 8 [0..0.14h, 0.02h 刻み] x 字高 4 = 1 字 96 試料、

40 字で 3840 試料)。

seed : 小数ずらし・隣字・雑音の乱数種(決定的)。

neighbour : "random"(既定)か字種内の 1 字。

Returns

dict

features (n, 88), labels (n,) str, weights (88,) 対角マハラノビスの重み,

blur (n,) 各試料のぼけ, blur_levels, blur_step, charset, font,

font_kind, k (=1), reject_distance(全体の値、帯を指定しない照合用),

reject_distance_by_class(字ごと), min_margin_ratio, reject_stats

(within_nn_p99, factor, between_nn_min, between_nn_p01), n_fallback,

heights。

棄却距離の導き方 / reject_distance derivation

各試料について「同クラス・ぼけ帯内(±1 格子点)の最近傍の他の試料までの重みつき距離」を

求め、字ごとにその 99 パーセンタイル x REJECT_FACTOR を棄却距離とする。

学習分布だけで決まる(係数の選び方は REJECT_FACTOR の注記を参照)。

Per class: REJECT_FACTOR x P99 of the within-class, within-band nearest-other-sample

weighted distance. Derived from the training distribution only.

Detailed usage guide

• symbol_reading family guide

References (sample data, literature)

• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).

• Operator provenance and references — the sources of the research/methods this op family came from.

• The canonical algorithm (author, year) and its uses are named in the family usage guide above.

Runnable examples (verified samples that actually call this op)

• poc_symbol_lot_ocr — py -3.11 examples/poc_symbol_lot_ocr.py

Ops the type connects to (they accept `table` as input)

render_symbol · symbol_read_trials · symbol_read_pod · ocr_classify · ocr_read_line · ocr_read_trials

Same category (`ocr`)

ocr_font · ocr_render_text · ocr_segment_line · ocr_features · ocr_classify · ocr_read_line · ocr_score · ocr_read_trials

*Provenance: ocrclassic.py — SYMBOL operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*

ocr_train — SYMBOL ocr op