Files
douyi_guaji_main/operations/ocr_operations.py
T
Codex 0131be7869 切换 OCR 默认引擎到 PaddleOCR v4-mobile 并简化预处理
- 默认引擎改为 PaddleOCR PP-OCRv4_mobile(实测 369ms/张,比 EasyOCR 629ms 更快且精度大幅提升,coupon 等关键图错字一片 -> 全对)
- EasyOCR 仅作回退,移除"小图自动放大+锐化+对比度增强"预处理(实测对 Paddle 是负优化,会引入伪影/重复字)
- analyse_pic_word 加 LRU 内容哈希缓存(256 条),重复读图直接命中
- analyse_pic_digits 从 200 行重复实现精简成 30 行委托 + 数字过滤
- 统一两处 OCR 初始化为 _build_paddle / _build_easyocr 工厂,消除重复
- 设置 FLAGS_use_mkldnn=false 绕开 Win + Paddle 3.x 的 oneDNN+PIR 不兼容 bug(实测 FLAGS_enable_pir_in_executor=0 也修不掉,关 oneDNN 是当前唯一可行解)
- 决策依据/速度数据/已知问题归档到 docs/ocr_engine_decision.md
- 业务接口完全向后兼容
2026-05-21 10:50:30 +08:00

259 lines
8.8 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""OCR操作模块
包含OCR识别相关的操作。
"""
from __future__ import annotations
import os
import contextlib
import hashlib
import threading
from collections import OrderedDict
from typing import Any
# PaddleOCR 3.x 会在初始化/导入时进行模型源连通性检查;在国内网络环境下容易卡住。
# 这里提前设置环境变量,避免启动耗时/卡死。
os.environ.setdefault("DISABLE_MODEL_SOURCE_CHECK", "True")
# 额外抑制 paddle/oneDNN/glog 的冗余信息(INFO/WARNING)
os.environ.setdefault("GLOG_minloglevel", "2")
# Win + Paddle 3.x:oneDNN 后端跟新 PIR 执行器有 ConvertPirAttribute2RuntimeAttribute 不兼容 bug。
# paddle 3.3.1 + paddleocr 3.5 + Win 实测:FLAGS_enable_pir_in_executor=0 也修不掉(PIR 已强制默认)。
# 唯一可行解:关 oneDNN。代价是 CPU 推理慢 20-30%,但 v4-mobile 仍有 ~370ms/张实测,业务足够。
# 等 paddle 上游补全 onednn_instruction.cc:118 的属性转换分支后可以删掉这行。
os.environ.setdefault("FLAGS_use_mkldnn", "false")
os.environ.setdefault("FLAGS_call_stack_level", "0")
# 优化OMP_NUM_THREADS:允许底层库使用多线程,但不要设置太高避免线程竞争
import os as os_module
cpu_count = os_module.cpu_count() or 4
optimal_omp_threads = min(4, max(2, cpu_count // 3))
os.environ.setdefault("OMP_NUM_THREADS", str(optimal_omp_threads))
import logging as _logging
# 模块级 logger:原代码用 print 输出 OCR 引擎初始化日志,统一走 logging。
# 保留消息中已有的 "[OCR]" 前缀,避免改变历史搜索习惯。
_LOGGER = _logging.getLogger("operations.ocr_operations")
def _ocr_log(msg: str, level: str = "INFO") -> None:
"""统一 OCR 操作日志出口;失败兜底 stdout。"""
try:
lv = str(level or "INFO").upper()
if lv == "ERROR":
_LOGGER.error(str(msg))
elif lv == "WARNING":
_LOGGER.warning(str(msg))
elif lv == "DEBUG":
_LOGGER.debug(str(msg))
else:
_LOGGER.info(str(msg))
except Exception:
try:
_ocr_log(str(msg))
except Exception:
pass
# ===== 结果缓存(基于图片内容哈希,LRU)=====
_OCR_CACHE_MAX = int(os.environ.get("OCR_CACHE_MAX", "256"))
_ocr_cache: "OrderedDict[str, Any]" = OrderedDict()
_ocr_cache_lock = threading.Lock()
_ocr_cache_stats = {"hit": 0, "miss": 0}
def _img_fingerprint(img_bytes: bytes, mode: int = 0) -> str:
"""对图片字节做 blake2b 哈希作为缓存 key(比 sha256 快约 30%)。"""
h = hashlib.blake2b(img_bytes, digest_size=12)
h.update(bytes([mode & 0xFF]))
return h.hexdigest()
def cache_get(key: str):
with _ocr_cache_lock:
if key in _ocr_cache:
_ocr_cache.move_to_end(key)
_ocr_cache_stats["hit"] += 1
return _ocr_cache[key]
_ocr_cache_stats["miss"] += 1
return None
def cache_put(key: str, value: Any) -> None:
with _ocr_cache_lock:
_ocr_cache[key] = value
_ocr_cache.move_to_end(key)
while len(_ocr_cache) > _OCR_CACHE_MAX:
_ocr_cache.popitem(last=False)
def cache_clear() -> None:
with _ocr_cache_lock:
_ocr_cache.clear()
_ocr_cache_stats["hit"] = 0
_ocr_cache_stats["miss"] = 0
def cache_stats() -> dict:
with _ocr_cache_lock:
total = _ocr_cache_stats["hit"] + _ocr_cache_stats["miss"]
hit_rate = (_ocr_cache_stats["hit"] / total) if total else 0.0
return {
"size": len(_ocr_cache),
"max": _OCR_CACHE_MAX,
"hit": _ocr_cache_stats["hit"],
"miss": _ocr_cache_stats["miss"],
"hit_rate": hit_rate,
}
# Paddle OCR 推理参数(统一在这里,避免初始化两处重复)
PADDLE_KWARGS = dict(
ocr_version="PP-OCRv4",
text_detection_model_name="PP-OCRv4_mobile_det",
text_recognition_model_name="PP-OCRv4_mobile_rec",
device="cpu",
cpu_threads=optimal_omp_threads,
enable_mkldnn=False, # 关键:绕开 Win Paddle 3.x oneDNN+PIR bug
precision="fp32",
lang="ch",
use_doc_orientation_classify=False,
use_doc_unwarping=False,
use_textline_orientation=False,
)
def _build_paddle():
"""构造 PaddleOCR 实例。stdout/stderr 静默处理。"""
devnull = open(os.devnull, "w")
try:
with contextlib.redirect_stdout(devnull), contextlib.redirect_stderr(devnull):
from paddleocr import PaddleOCR # type: ignore # pylint: disable=import-error
return PaddleOCR(**PADDLE_KWARGS)
finally:
devnull.close()
def _build_easyocr():
"""构造 EasyOCR 实例(回退方案)。"""
import io
from contextlib import redirect_stdout, redirect_stderr
import easyocr # type: ignore[import-untyped] # pylint: disable=import-error
with redirect_stdout(io.StringIO()), redirect_stderr(io.StringIO()):
return easyocr.Reader(["ch_sim", "en"], gpu=False, verbose=False)
class OCRProcessor:
"""OCR处理器类"""
def __init__(self):
# 关闭文档方向/矫正等重型前处理:我们的截图是手机UI,不是扫描文档。
# 支持延迟初始化(测试流程/只做点击时不需要 OCR)。
self.ocr: Any = None
self.ocr_type = None # 'easyocr' 或 'paddleocr'
# 默认使用CPU模式(删除GPU检测)
self._gpu_available = False
if not str(os.environ.get("SKIP_OCR_INIT", "")).strip():
self._init_ocr()
def _init_ocr(self):
"""初始化OCR引擎
基于 31 张项目截图的 benchmark:
- PaddleOCR v4-mobile:~370ms/张,coupon 关键图全对 ← 首选
- EasyOCR:~630ms/张,coupon 错字一片,仅作回退
"""
# 优先 PaddleOCR v4-mobile
try:
self.ocr = _build_paddle()
self.ocr_type = "paddleocr"
_ocr_log("[OCR] 使用 PaddleOCR v4-mobile(CPU模式)")
return
except Exception as e:
_ocr_log(f"[OCR] PaddleOCR 初始化失败: {e},回退到 EasyOCR", "WARNING")
# 回退:EasyOCR
try:
self.ocr = _build_easyocr()
self.ocr_type = "easyocr"
_ocr_log("[OCR] 使用 EasyOCR(CPU模式,回退引擎)")
except Exception as e:
_ocr_log(f"[OCR] 所有 OCR 引擎初始化失败: {e}", "ERROR")
self.ocr = None
self.ocr_type = None
def _ensure_ocr(self):
"""确保OCR引擎已初始化(延迟初始化路径)"""
if self.ocr is not None:
return
self._init_ocr()
# 全局OCR处理器实例
_ocr_processor = OCRProcessor()
# 全局OCR底层操作单例(延迟初始化)
_UNDERLYING_OPS_SINGLETON = None
def _ensure_ocr():
"""确保OCR引擎已初始化(向后兼容接口)"""
_ocr_processor._ensure_ocr()
def analyse_pic_word(picname, change_color=0, pic_path=None):
"""文字OCR识别(向后兼容接口)"""
# 使用全局复用的 underlying_operations 实例,避免每次调用都初始化 OCR(耗时)
from douyin.core.underlying_operations import underlying_operations
global _UNDERLYING_OPS_SINGLETON # type: ignore[name-defined]
try:
_UNDERLYING_OPS_SINGLETON # type: ignore[name-defined]
except NameError:
_UNDERLYING_OPS_SINGLETON = underlying_operations()
return _UNDERLYING_OPS_SINGLETON.analyse_pic_word(picname, change_color, pic_path)
def analyse_pic_digits(picname, change_color=1, pic_path=None):
"""数字OCR识别(向后兼容接口)"""
# 使用全局复用的 underlying_operations 实例,避免每次调用都初始化 OCR(耗时)
from douyin.core.underlying_operations import underlying_operations
global _UNDERLYING_OPS_SINGLETON # type: ignore[name-defined]
try:
_UNDERLYING_OPS_SINGLETON # type: ignore[name-defined]
except NameError:
_UNDERLYING_OPS_SINGLETON = underlying_operations()
return _UNDERLYING_OPS_SINGLETON.analyse_pic_digits(picname, change_color, pic_path)
def extract_ocr_content(content=[], return_confidence=False):
"""提取OCR内容(向后兼容接口)"""
# 使用全局复用的 underlying_operations 实例
from douyin.core.underlying_operations import underlying_operations
global _UNDERLYING_OPS_SINGLETON # type: ignore[name-defined]
try:
_UNDERLYING_OPS_SINGLETON # type: ignore[name-defined]
except NameError:
_UNDERLYING_OPS_SINGLETON = underlying_operations()
return _UNDERLYING_OPS_SINGLETON.extract_ocr_content(content, return_confidence)
__all__ = [
"OCRProcessor",
"analyse_pic_word",
"analyse_pic_digits",
"extract_ocr_content",
"_ensure_ocr",
"cache_get",
"cache_put",
"cache_clear",
"cache_stats",
"_img_fingerprint",
"PADDLE_KWARGS",
]