python pdf 转 md 文档 多个方案

推荐

常规pdf 推荐方案2，方案1

pdf中 图片多 扫描件 推荐ocr 方案3 

有大模型基础和资源 推荐 方案4 

方案1 pdf转word转html转md

解析较慢 链路较多 效果和 方案2类似 但不如方案2快

import time
import mammoth
import markdownify
from pdf2docx import Converter
# 参考 https://zhuanlan.zhihu.com/p/688159597  https://cloud.tencent.com/developer/article/1768983


# 转存Word文档内的图片
def convert_images(image):
    with image.open() as image_bytes:
        file_suffix = image.content_type.split("/")[1]
        path_file = base_path+"{}.{}".format(str(time.time()),file_suffix)
        with open(path_file, 'wb') as f:
            f.write(image_bytes.read())

    return {"src":path_file}


def pdf2md(pdf_path):
    start = time.time()
    word_path = pdf_path+'.docx'
    html_path = pdf_path+'.docx.html'
    md_path = pdf_path+'.docx.html.md'
    # 创建对象
    cv = Converter(pdf_path)
    # 实现 PDF 转换为 Word
    cv.convert(word_path, start=0, end=None)
    # 关闭对象
    cv.close()
    # 转化Word文档为HTML
    result = mammoth.convert_to_html(word_path,convert_image = mammoth.images.img_element(convert_images))
    # 获取HTML内容
    html = result.value
    # 转化HTML为Markdown
    md = markdownify.markdownify(html,heading_style="ATX")
    with open(html_path,'w',encoding='utf-8') as html_file,open(md_path,"w",encoding='utf-8') as md_file:
        html_file.write(html)
        md_file.write(md)

    elapse_time = time.time() - start
    print('转换 '+pdf_path+' 运行时间：'+ str(elapse_time))



if __name__ == '__main__':
    # pdf转doc转html转md
    base_path = '/Users/xxxx/'
    pdf2md(base_path+'xxxxx.pdf')
 

方案2 使用 pymupdf 的  pymupdf4llm

解析快 会存在字体缺失，乱码问题

import time
import pymupdf4llm

# 参考 https://pymupdf.readthedocs.io/en/latest/rag.html

# pdf

def pdf2md(pdf_path,md_path):
    start = time.time()
    md_text = pymupdf4llm.to_markdown(pdf_path, write_images=True)
    output = open(md_path, "w")
    output.write(md_text)
    output.close()
    elapse_time = time.time() - start
    print('转换 '+pdf_path+' 运行时间：'+ str(elapse_time))


if __name__ == '__main__':
    base_path = '/Users/xxxx/'
    pdf2md(base_path+'xxxxx.pdf',base_path+'xxxx.pdf.md')

方案3 使用 pix2text 进行 ocr识别 

pix2text 安装会很大，效果不错，解析时间会较长

import time
from pix2text import Pix2Text

# 主要使用ocr识别 参考 https://www.breezedeus.com/article/p2t-v1.1
# 类似项目 VikParuchuri/marker  https://github.com/VikParuchuri/marker?tab=readme-ov-file


def pdf2md(base_path,pdf_path,md_name):
    start = time.time()
    p2t = Pix2Text.from_config(enable_formula=False)
    doc = p2t.recognize_pdf(pdf_path)
    doc.to_markdown(out_dir=base_path,markdown_fn=md_name)
    elapse_time = time.time() - start
    print('转换 '+pdf_path+' 运行时间：'+ str(elapse_time))


if __name__ == '__main__':
    base_path = '/Users/xxx/'
    pdf2md(base_path,base_path+'xxx.pdf','xxx.pdf.md')

方案4 使用 ai模型 gpt 问答

效果不可控，解析时间长，需要微调和适配

参考 https://mp.weixin.qq.com/s/P5KBcqXBqHqvrqnlMIkQYA