"""ONNX-adjunct Tesseract beam producer for whitelisted numeric crops.""" from __future__ import annotations from .arabic_tesseract_engine import ( POLARITY_RETRY_MEAN_CONF, bytes_to_bgr, extract_structured_numeric_tokens, invert_polarity, mean_token_confidence, noir_filter_preprocess, tokens_to_text, ) from .contracts import BeamCandidate, SemanticClass def _tokens_to_beam( tokens: list[dict], text: str, variant: str, ) -> BeamCandidate | None: if not text: return None confidences = [max(0.0, min(1.0, float(item["confidence"]) / 100.0)) for item in tokens] if not confidences: confidences = [0.5] * len(text) elif len(confidences) < len(text): mean = sum(confidences) / len(confidences) confidences = confidences + [mean] * (len(text) - len(confidences)) else: confidences = confidences[: len(text)] mean_prob = sum(confidences) / len(confidences) min_prob = min(confidences) return BeamCandidate( text=text, probability=mean_prob, min_character_probability=min_prob, beam_margin=0.15, character_confidences=tuple(confidences), character_confidence_sources=tuple("engine" for _ in confidences), preprocessing_variant=variant, ) class TesseractNumericRecognizer: version = "tesseract-numeric-v1" @staticmethod def available() -> bool: try: import cv2 # noqa: F401 import pytesseract pytesseract.get_tesseract_version() return True except Exception: return False def recognize(self, image_bytes: bytes, semantic_class: SemanticClass) -> list[BeamCandidate]: if not self.available(): return [] try: bgr = bytes_to_bgr(image_bytes) except ValueError: return [] beams: list[BeamCandidate] = [] processed = noir_filter_preprocess(bgr) tokens = extract_structured_numeric_tokens( bgr, semantic_class=semantic_class, processed=processed ) text = tokens_to_text(tokens, semantic_class) primary = _tokens_to_beam(tokens, text, "tesseract_adaptive") if primary: beams.append(primary) if mean_token_confidence(tokens) < POLARITY_RETRY_MEAN_CONF: inverted = invert_polarity(processed) retry_tokens = extract_structured_numeric_tokens( bgr, semantic_class=semantic_class, processed=inverted ) retry_text = tokens_to_text(retry_tokens, semantic_class) retry = _tokens_to_beam(retry_tokens, retry_text, "tesseract_polarity_inverted") if retry and retry.text not in {item.text for item in beams}: beams.append(retry) return beams