ovkit.recognize.ocr의 소스 코드

"""Text-recognition adapter — greedy CTC decoding.

OMZ text-recognition models (``text-recognition-*``, ``handwritten-*``,
``license-plate-recognition-*``) emit per-timestep symbol logits, typically
``[T, 1, C]`` (or ``[1, T, C]``). Greedy CTC: argmax per step, collapse repeats,
drop blanks. The decoded string is stored in :attr:`Results.text` and the raw
logits stay available in :attr:`Results.tensors`.

The symbol set defaults to OMZ ``text-recognition-0012``'s alphabet
(``0-9a-z#`` with ``#`` as blank); override per model via the manifest
``postprocess.symbols`` / ``postprocess.blank`` fields.

A Korean alphabet has ~3,700 symbols, which is a data file and not a manifest
string, so ``postprocess.charset: labels`` reads the table from the
``labels.txt`` that travels with the model — one symbol per line. PaddleOCR
recognisers also number their classes the other way round (blank **first**,
so dictionary entry *i* is class *i+1*) and append the space character in
code rather than in the file: ``blank_first`` and ``space`` say so.
"""

from __future__ import annotations

from typing import Any

import numpy as np

from ..core.backend import Backend
from ..core.results import Results
from .base import BaseAdapter

#: Default symbol table (OMZ text-recognition-0012): blank ('#') last.
_DEFAULT_SYMBOLS = "0123456789abcdefghijklmnopqrstuvwxyz#"

#: A space cannot be a line in a labels file — anything that reads one strips
#: it away — so a charset writes it as this token and the decoder maps it back.
#: Dropping it instead would shift every later class by one.
_SPACE_TOKEN = "<space>"




class OCRAdapter(BaseAdapter):
    """Adapter for text recognition (greedy CTC)."""

    task = "optical_character_recognition"



    def run(self, backend: Backend, image: np.ndarray, **_: Any) -> Results:
        size = self.model_input_hw(backend)
        rgb = bool(self.pre.get("rgb", False))
        feed = self.preprocess(image, size, rgb=rgb, scale=self.pre.get("scale", 1.0))
        outputs = backend.infer(feed)

        logits = np.asarray(next(iter(outputs.values())), dtype=np.float32)
        text = self._ctc_greedy(logits)
        res = Results(image, task=self.task, names=self.names, tensors=outputs)
        res.text = text
        return res


    def _symbols(self) -> tuple[list[str], int]:
        """The symbol table and which class id means "nothing here"."""
        if self.post.get("charset") == "labels" and self.names:
            table = [
                " " if self.names[i] == _SPACE_TOKEN else self.names[i] for i in sorted(self.names)
            ]
            if self.post.get("space"):
                table.append(" ")
            if self.post.get("blank_first"):
                # Class 0 is the blank; dictionary entry i is class i + 1.
                return [""] + table, 0
            return table, len(table) - 1

        table = list(str(self.post.get("symbols", _DEFAULT_SYMBOLS)))
        blank = self.post.get("blank")
        return table, int(blank) if blank is not None else len(table) - 1

    def _ctc_greedy(self, logits: np.ndarray) -> str:
        """Greedy CTC decode of ``[T, 1, C]`` / ``[1, T, C]`` / ``[T, C]`` logits."""
        a = logits
        if a.ndim == 3:  # [T,1,C] or [1,T,C] -> [T,C]
            a = a[:, 0, :] if a.shape[1] == 1 else a[0]
        if a.ndim != 2:
            return ""

        symbols, blank_idx = self._symbols()

        ids = a.argmax(axis=1)
        chars: list[str] = []
        prev = -1
        for i in ids:
            i = int(i)
            if i != prev and i != blank_idx and i < len(symbols):
                chars.append(symbols[i])
            prev = i
        return "".join(chars)