【maker-pdf 文档文字识别（包含ocr），安装使用完整教程】

测试效果还比较好，比markitdown要好

安装环境

conda create -n maker-pdf python=3.12
conda activate marker-pdf
pip install modelscope
pip install marker-pdf -U

下载模型

• 建议用modelscope上缓存的模型，不然下载会很慢

from modelscope import snapshot_download

model_root = "models"
snapshot_download("Lixiang/marker-pdf", local_dir="models")

开始运行

• 初始化

from marker.converters.pdf import PdfConverter
from marker.models import create_model_dict
from marker.output import text_from_rendered
from surya.settings import settings

# 如果想用官方最新的权重，请注释掉下边的代码-------------------------
## 权重地址，模型很大，没GPU的话会很慢
model_root = "models"
settings.MODEL_CACHE_DIR = model_root
for chectpoint in [
    "LAYOUT_MODEL_CHECKPOINT",
    "DETECTOR_MODEL_CHECKPOINT",
    "OCR_ERROR_MODEL_CHECKPOINT",
    "TABLE_REC_MODEL_CHECKPOINT",
    "RECOGNITION_MODEL_CHECKPOINT",
]:
    value = getattr(settings, chectpoint)
    if "s3://" in value:
	    value = value.replace("s3://", "/")
	    setattr(settings, chectpoint, model_root + value)
 # -------------------------------------------------------------

converter = PdfConverter(
    artifact_dict=create_model_dict(),
)

• 推理pdf

rendered = converter("test.pdf")
# text = rendered.markdown
text, _, images = text_from_rendered(rendered)
print(text)