163 lines
5.2 KiB
Python
163 lines
5.2 KiB
Python
from __future__ import annotations
|
|||
|
|
|
||
|
|
import threading
|
||
|
|
from pathlib import Path
|
||
|
|
from typing import Callable
|
||
|
|
|
||
|
|
from loguru import logger
|
||
|
|
from sqlalchemy import select
|
||
|
|
from sqlalchemy.orm import Session
|
||
|
|
|
||
|
|
from sakura_cloud.core.config import settings
|
||
|
|
from sakura_cloud.db.session import SessionLocal
|
||
|
|
from sakura_cloud.models.artwork import Artwork
|
||
|
|
from sakura_cloud.models.ocr import ArtworkOcrText
|
||
|
|
|
||
|
|
|
||
|
|
OcrExtractor = Callable[[Path], tuple[str, float | None, str]]
|
||
|
|
_semaphore = threading.BoundedSemaphore(settings.ocr_max_concurrency)
|
||
|
|
|
||
|
|
|
||
|
|
def queue_ocr_for_artwork(db: Session, artwork: Artwork) -> ArtworkOcrText | None:
|
||
|
|
if not settings.ocr_enabled:
|
||
|
|
return None
|
||
|
|
row = _get_or_create_row(db, artwork)
|
||
|
|
if row.status in {"pending", "running", "done"}:
|
||
|
|
return row
|
||
|
|
row.status = "pending"
|
||
|
|
row.error = None
|
||
|
|
db.commit()
|
||
|
|
return row
|
||
|
|
|
||
|
|
|
||
|
|
def run_ocr_for_artwork_id(artwork_id: str, user_id: str) -> None:
|
||
|
|
with SessionLocal() as db:
|
||
|
|
artwork = db.get(Artwork, artwork_id)
|
||
|
|
if artwork is None or artwork.user_id != user_id:
|
||
|
|
return
|
||
|
|
run_ocr_for_artwork(db, artwork)
|
||
|
|
|
||
|
|
|
||
|
|
def run_ocr_for_artwork(
|
||
|
|
db: Session,
|
||
|
|
artwork: Artwork,
|
||
|
|
*,
|
||
|
|
extractor: OcrExtractor | None = None,
|
||
|
|
) -> ArtworkOcrText:
|
||
|
|
row = _get_or_create_row(db, artwork)
|
||
|
|
if not settings.ocr_enabled and extractor is None:
|
||
|
|
row.status = "failed"
|
||
|
|
row.error = "ocr_disabled"
|
||
|
|
db.commit()
|
||
|
|
return row
|
||
|
|
if not _is_supported_image(artwork):
|
||
|
|
row.status = "failed"
|
||
|
|
row.error = "unsupported_file_type"
|
||
|
|
db.commit()
|
||
|
|
logger.info("OCR skipped artwork={} mime={}", artwork.id, artwork.mime_type)
|
||
|
|
return row
|
||
|
|
if row.status == "running":
|
||
|
|
return row
|
||
|
|
if row.attempts >= settings.ocr_max_attempts and extractor is None:
|
||
|
|
return row
|
||
|
|
|
||
|
|
acquired = _semaphore.acquire(blocking=False)
|
||
|
|
if not acquired:
|
||
|
|
row.status = "pending"
|
||
|
|
row.error = "ocr_busy"
|
||
|
|
db.commit()
|
||
|
|
return row
|
||
|
|
try:
|
||
|
|
row.status = "running"
|
||
|
|
row.attempts += 1
|
||
|
|
row.error = None
|
||
|
|
db.commit()
|
||
|
|
|
||
|
|
text, confidence, language = (extractor or _paddle_ocr_extract)(Path(artwork.file_path))
|
||
|
|
row.ocr_text = text.strip()
|
||
|
|
row.confidence = confidence
|
||
|
|
row.language = language or "mixed"
|
||
|
|
row.engine = "paddleocr"
|
||
|
|
row.status = "done"
|
||
|
|
row.error = None
|
||
|
|
db.commit()
|
||
|
|
logger.info("OCR done artwork={} text_len={} confidence={}", artwork.id, len(row.ocr_text), confidence)
|
||
|
|
return row
|
||
|
|
except Exception as exc:
|
||
|
|
row.status = "failed"
|
||
|
|
row.error = exc.__class__.__name__
|
||
|
|
db.commit()
|
||
|
|
logger.warning("OCR failed artwork={} attempt={} error={}", artwork.id, row.attempts, exc.__class__.__name__)
|
||
|
|
return row
|
||
|
|
finally:
|
||
|
|
_semaphore.release()
|
||
|
|
|
||
|
|
|
||
|
|
def _get_or_create_row(db: Session, artwork: Artwork) -> ArtworkOcrText:
|
||
|
|
row = db.scalar(
|
||
|
|
select(ArtworkOcrText).where(
|
||
|
|
ArtworkOcrText.user_id == artwork.user_id,
|
||
|
|
ArtworkOcrText.artwork_id == artwork.id,
|
||
|
|
)
|
||
|
|
)
|
||
|
|
if row is None:
|
||
|
|
row = ArtworkOcrText(
|
||
|
|
user_id=artwork.user_id,
|
||
|
|
artwork_id=artwork.id,
|
||
|
|
file_id=artwork.file_id,
|
||
|
|
source_platform=artwork.source_platform,
|
||
|
|
source_id=artwork.source_id,
|
||
|
|
status="pending",
|
||
|
|
)
|
||
|
|
db.add(row)
|
||
|
|
db.commit()
|
||
|
|
db.refresh(row)
|
||
|
|
else:
|
||
|
|
row.file_id = artwork.file_id
|
||
|
|
row.source_platform = artwork.source_platform
|
||
|
|
row.source_id = artwork.source_id
|
||
|
|
return row
|
||
|
|
|
||
|
|
|
||
|
|
def _is_supported_image(artwork: Artwork) -> bool:
|
||
|
|
mime = (artwork.mime_type or "").lower()
|
||
|
|
if mime in {"image/jpeg", "image/png", "image/webp"}:
|
||
|
|
return True
|
||
|
|
suffix = Path(artwork.file_path or "").suffix.lower()
|
||
|
|
return suffix in {".jpg", ".jpeg", ".png", ".webp"} and mime != "image/gif"
|
||
|
|
|
||
|
|
|
||
|
|
def _paddle_ocr_extract(path: Path) -> tuple[str, float | None, str]:
|
||
|
|
from PIL import Image
|
||
|
|
from paddleocr import PaddleOCR
|
||
|
|
|
||
|
|
if not path.is_file():
|
||
|
|
raise FileNotFoundError("image_not_found")
|
||
|
|
|
||
|
|
work_path = path
|
||
|
|
with Image.open(path) as image:
|
||
|
|
original_size = image.size
|
||
|
|
image.thumbnail((settings.ocr_max_image_dimension, settings.ocr_max_image_dimension))
|
||
|
|
if image.size != original_size:
|
||
|
|
tmp = path.with_suffix(".ocr-resized.jpg")
|
||
|
|
image.convert("RGB").save(tmp, quality=90)
|
||
|
|
work_path = tmp
|
||
|
|
|
||
|
|
ocr = PaddleOCR(use_angle_cls=True, lang=settings.ocr_language, use_gpu=False)
|
||
|
|
result = ocr.ocr(str(work_path), cls=True)
|
||
|
|
texts: list[str] = []
|
||
|
|
scores: list[float] = []
|
||
|
|
for page in result or []:
|
||
|
|
for line in page or []:
|
||
|
|
if len(line) >= 2 and isinstance(line[1], (list, tuple)):
|
||
|
|
text = str(line[1][0]).strip()
|
||
|
|
if text:
|
||
|
|
texts.append(text)
|
||
|
|
try:
|
||
|
|
scores.append(float(line[1][1]))
|
||
|
|
except (TypeError, ValueError, IndexError):
|
||
|
|
pass
|
||
|
|
if work_path != path:
|
||
|
|
work_path.unlink(missing_ok=True)
|
||
|
|
confidence = sum(scores) / len(scores) if scores else None
|
||
|
|
return "\n".join(texts), confidence, settings.ocr_language
|