ai之pdf解析工具 PPStructure 还是PaddleOCR

重点是四 先用 PPStructure 版面分析，分成不同的块儿，再选用 PaddleOCR、或PPStructure

https://chat.deepseek.com/a/chat/s/1dafb23d-d230-48a6-b3a1-72569eb1c7cd

我有解析PDF的程序如下：
1、配置
“”"
OCR处理工具类
功能：统一管理OCR引擎的初始化、处理、资源回收
“”"
import multiprocessing
import os
from pathlib import Path
from typing import Tuple, Optional, Dict, Any

import paddle
from paddleocr import PaddleOCR, PPStructure

from .GPUResource import GPUResourceController
from …utils.logutil import get_local_logger

logger = get_local_logger()

基础路径

BASE_DIR = Path(file).parent.parent
MODEL_DIR = BASE_DIR / “common/models”
logger.info(f" >>>>>> MODEL_DIR路径: {MODEL_DIR}")

OCR模型配置

START_PAGE = 6

OCR模型配置

OCR_CONFIG = {
# 显式指定所有模型路径
# det_model_dir=str(MODEL_DIR /“ch_PP-OCRv4_det_infer”),
“det_model_dir”: str(MODEL_DIR / “ch_PP-OCRv4_det_infer”),
“rec_model_dir”: str(MODEL_DIR / “ch_PP-OCRv4_rec_infer”),
“cls_model_dir”: str(MODEL_DIR / “ch_ppocr_mobile_v2.0_cls_infer”),
“table_model_dir”: str(MODEL_DIR / “en_ppstructure_mobile_v2.0_table_structure_infer”),
“layout_model_dir”: str(MODEL_DIR / “en_ppstructure_mobile_v2.0_layout_infer”),
“lang”: “ch”,
“ocr_version”: “PP-OCRv4”,
“table_version”: “PP-StructureV2”
}

GPU配置

GPU_CONFIG = {
“enable_gpu”: True,
“max_gpus”: 2,
“memory_fraction”: 0.5
}

硬件配置

GPU_MEMORY_LIMIT = 0.8 # 单进程最大显存占比
CPU_THREADS = 4 # CPU模式线程数

性能配置

PRECISION_MODE = “fp16” if paddle.is_compiled_with_cuda() else “fp32”
ENABLE_TENSORRT = True

use_gpu_flag = paddle.device.is_compiled_with_cuda()
print(f"torch.cuda.is_available(){use_gpu_flag}")
print(f"torch.version{paddle.device.get_cudnn_version()}")

class OCREngineInitError(Exception):
“”“OCR引擎初始化异常”“”
pass

“”"
OCR配置工具类
“”"
class OCRConfig:

def __init__(self, process_idx: int = 0):"""初始化OCR工具:param process_idx: 进程ID（用于多进程GPU分配）"""self.process_idx = process_idxself.gpu_id = -1self.ocr_engine = Noneself.table_engine = Noneself._init_gpu()self._init_ocr_engine()self._init_table_engine()def __enter__(self):self._set_hardware_environment()self._init_engines()return selfdef __exit__(self, exc_type, *_):self._cleanup_resources()self._report_gpu_status(exc_type is None)def _set_hardware_environment(self):os.environ["CUDA_VISIBLE_DEVICES"] = str(self.gpu_id) if self.gpu_id != -1 else ""device = "gpu:0" if self.gpu_id != -1 else "cpu"paddle.set_device(device)if self.gpu_id != -1:paddle.set_flags({'FLAGS_fraction_of_gpu_memory_to_use': GPU_MEMORY_LIMIT,'FLAGS_allocator_strategy': 'auto_growth'})def _init_gpu(self) -> None:"""初始化GPU环境"""if not GPU_CONFIG["enable_gpu"] or not paddle.is_compiled_with_cuda():logger.info(f" paddle.is_compiled_with_cuda: {paddle.device.is_compiled_with_cuda()}, 即将运行在CPU模式")paddle.set_device("cpu")returnnum_gpus = paddle.device.cuda.device_count()if num_gpus == 0:logger.warn(f" num_gpus: {num_gpus}, 没有检测到可用的GUP, 即将运行在CPU模式")return -1  # 标记为CPU模式self.gpu_id = self.process_idx % num_gpusos.environ["CUDA_VISIBLE_DEVICES"] = str(self.gpu_id)paddle.set_device(f"gpu:{self.gpu_id}")logger.info(f"初始化GPU:{self.gpu_id} | 进程:{self.process_idx}")# 设置显存分配比例paddle.set_flags({'FLAGS_fraction_of_gpu_memory_to_use': GPU_CONFIG["memory_fraction"]})def _init_ocr_engine(self) -> None:"""初始化OCR引擎"""try:self.ocr_engine = PaddleOCR(use_angle_cls=True,lang=OCR_CONFIG["lang"],page_num=START_PAGE,use_gpu=self.gpu_id != -1,gpu_id=self.gpu_id,table_model_dir="../common/models/en_ppstructure_mobile_v2.0_table_structure_infer",  # 显式指定表格模型# 显式指定所有模型路径# det_model_dir=str(MODEL_DIR /"ch_PP-OCRv4_det_infer"),det_model_dir=OCR_CONFIG["det_model_dir"],rec_model_dir=OCR_CONFIG["rec_model_dir"],cls_model_dir=OCR_CONFIG["cls_model_dir"],# ocr_version=OCR_CONFIG["ocr_version"],table_version="PP-StructureV2",  # 使用表格识别模型show_log=False,# 性能优化参数enable_mkldnn=True,  # Intel CPU 加速use_tensorrt=False,   # NVIDIA GPU 加速precision="fp16"     # 混合精度推理)print(f" OCR引擎初始化成功 [PID:{os.getpid()}] Assigned to GPU {self.gpu_id} (Process index: {self.process_idx})")except Exception as e:logger.error(f"OCR引擎初始化失败: {str(e)}")raise OCREngineInitError("OCR初始化失败") from edef _init_table_engine(self) -> None:"""初始化表格引擎: 使用RapidTable """# table_input = RapidTableInput(#     model_type="unitable",#     use_cuda=self.gpu_id != -1,#     device=f"cuda:{self.gpu_id}" if self.gpu_id != -1 else "cpu"# )# self.table_engine = RapidTable(table_input)"""初始化表格引擎: 使用PPStructure """try:self.table_engine = PPStructure(table_model_dir=OCR_CONFIG["table_model_dir"],layout_model_dir=OCR_CONFIG["layout_model_dir"],table_version=OCR_CONFIG["table_version"],ocr=False,use_gpu=GPU_CONFIG["enable_gpu"],gpu_id=self.gpu_id,show_log=False)logger.info("表格引擎初始化成功")except Exception as e:logger.error(f"表格引擎初始化失败: {str(e)}")raise OCREngineInitError("表格引擎初始化失败") from edef _cleanup_resources(self):"""资源清理"""if self.gpu_id != -1:paddle.device.cuda.empty_cache()logger.info(f"已释放GPU:{self.gpu_id}资源")
def _report_gpu_status(self, success: bool):if self.gpu_id != -1:GPUResourceController().release_gpu(self.gpu_id, success)

class OCRManager:
“”"
OCR引擎管理器（多进程安全）
“”"
_engines = {}

@classmethod
def get_engine(cls, process_idx: int) -> OCRConfig:"""获取进程专用引擎"""if not process_idx:process_idx = os.getpid()if process_idx not in cls._engines:try:cls._engines[process_idx] = OCRConfig(process_idx)except OCREngineInitError as e:raise RuntimeError(f"无法为进程{process_idx}创建引擎") from ereturn cls._engines[process_idx]
"""
GPU资源管理器（进程安全版本）
核心功能：
- 自动检测可用GPU
- 实现进程级GPU分配
- 避免跨进程污染
"""
_gpu_ring = []
_lock = multiprocessing.Lock()def __init__(self):self._init_gpu_pool()def _init_gpu_pool(self):"""安全初始化GPU资源池"""with self._lock:if not self._gpu_ring:if paddle.is_compiled_with_cuda():num_gpus = paddle.device.cuda.device_count()self._gpu_ring = list(range(num_gpus)) if num_gpus > 0 else []logger.info(f"GPU资源池初始化完成: {self._gpu_ring}")def acquire_gpu(self, process_id: int) -> int:"""进程安全获取GPU ID"""with self._lock:if not self._gpu_ring:return -1return self._gpu_ring[process_id % len(self._gpu_ring)]

2、处理程序：
def _process_ocr_and_table(self, img_path, page_idx:int):
# 假设这是你的 OCR 和表格识别逻辑
try:
“”“使用进程级引擎处理”“”
ocr_config = OCRManager.get_engine(os.getpid())
ocr_engine = ocr_config.ocr_engine
table_engine = ocr_config.table_engine

        # OCR识别ocr_output = ocr_engine.ocr(img_path)if not ocr_output:raise ValueError(" ocr_output OCR未识别到内容")logger.info(f" 第{page_idx}页OCR识别结果，类型：{type(ocr_output)}, \n 内容示例：{ocr_output}")# 解析OCR结果：每个元素的结构为 [box_coords, (text, score)]boxes = [line[0] for line in ocr_output]  # 提取所有框坐标txts = [line[1][0] for line in ocr_output]  # 提取所有文本