单个单词的 PDFMiner 提取 - LTText LTTextBox

Question

我在下面的示例中使用 PDFMiner 生成单词 x,y 坐标，但是结果是逐行生成的。我怎样才能将每个单词从另一个单词中拆分出来，而不是逐行拆分单词组（参见下面的示例）。我已经尝试了 PDFMiner tutorial 中的几个参数。 LTTextBox 和 LTText 都试过了。此外，我不能使用文本分析中通常使用的开始和结束偏移量。

这个 PDF 是一个很好的例子，它用在下面的代码中。

http://www.africau.edu/images/default/sample.pdf

from pdfminer.layout import LAParams, LTTextBox, LTText
from pdfminer.pdfpage import PDFPage
from pdfminer.pdfinterp import PDFPageInterpreter, PDFResourceManager
from pdfminer.converter import PDFPageAggregator

#Imports Searchable PDFs and prints x,y coordinates
fp = open('C:\sample.pdf', 'rb')
manager = PDFResourceManager()
laparams = LAParams()
dev = PDFPageAggregator(manager, laparams=laparams)
interpreter = PDFPageInterpreter(manager, dev)
pages = PDFPage.get_pages(fp)

for page in pages:
    print('--- Processing ---')
    interpreter.process_page(page)
    layout = dev.get_result()
    for lobj in layout:
        if isinstance(lobj, LTText):
            x, y, text = lobj.bbox[0], lobj.bbox[3], lobj.get_text()
            print('At %r is text: %s' % ((x, y), text))

此 returns 可搜索 PDF 的 x,y 坐标如下所示：

--- Processing ---
At (57.375, 747.903) is text: A Simple PDF File
At (69.25, 698.098) is text: This is a small demonstration .pdf file -
At (69.25, 674.194) is text: just for use in the Virtual Mechanics tutorials. More text. And more 
 text. And more text. And more text. And more text.

想要的结果（坐标代为演示）：

--- Processing ---
At (57.375, 747.903) is text: A
At (69.25, 698.098) is text: Simple
At (69.25, 674.194) is text: PDF
At (69.25, 638.338) is text: File

Answer 1

使用 PDFMiner，在遍历每一行之后（就像你已经做的那样），你只能遍历行中的每个字符。

我用下面的代码做到了这一点，同时尝试记录每个单词第一个字符的 x、y 并设置条件以在每个 LTAnno（例如 \n ）或 .get_text() == ' '空space.

from pdfminer.layout import LAParams, LTTextBox, LTText, LTChar, LTAnno
from pdfminer.pdfpage import PDFPage
from pdfminer.pdfinterp import PDFPageInterpreter, PDFResourceManager
from pdfminer.converter import PDFPageAggregator

#Imports Searchable PDFs and prints x,y coordinates
fp = open('C:\sample.pdf', 'rb')
manager = PDFResourceManager()
laparams = LAParams()
dev = PDFPageAggregator(manager, laparams=laparams)
interpreter = PDFPageInterpreter(manager, dev)
pages = PDFPage.get_pages(fp)

for page in pages:
    print('--- Processing ---')
    interpreter.process_page(page)
    layout = dev.get_result()
    x, y, text = -1, -1, ''
    for textbox in layout:
        if isinstance(textbox, LTText):
          for line in textbox:
            for char in line:
              # If the char is a line-break or an empty space, the word is complete
              if isinstance(char, LTAnno) or char.get_text() == ' ':
                if x != -1:
                  print('At %r is text: %s' % ((x, y), text))
                x, y, text = -1, -1, ''     
              elif isinstance(char, LTChar):
                text += char.get_text()
                if x == -1:
                  x, y, = char.bbox[0], char.bbox[3]    
    # If the last symbol in the PDF was neither an empty space nor a LTAnno, print the word here
    if x != -1:
      print('At %r is text: %s' % ((x, y), text))

输出如下所示

At (64.881, 747.903) is text: A
At (90.396, 747.903) is text: Simple
At (180.414, 747.903) is text: PDF
At (241.92, 747.903) is text: File

也许您可以优化条件来检测您的要求和喜好的词。（例如，在词尾剪掉标点符号 .!?）

Answer 2

读者可能想尝试的另一个包是 pdfparser，它也是基于 Poppler 构建的（使用 Cyton 绑定）并且恰好在性能上得到了更优化

                           pdfreader      pdfminer  speed-up factor

tiny document (half page)    0.033s         0.121s     3.6 x
small document (5 pages)     0.141s         0.810s     5.7 x
medium document (55 pages)   1.166s        10.524s     9.0 x
large document (436 pages)  10.581s       108.095s    10.2 x

除了速度更快之外，它的错误处理也更好，并解决了 Pdfminer 遇到的几个问题

import pdfparser.poppler as pdf
import sys

d=pdf.Document(sys.argv[1])

print('No of pages', d.no_of_pages)
for p in d:
    print('Page', p.page_no, 'size =', p.size)
    for f in p:
        print(' '*1,'Flow')
        for b in f:
            print(' '*2,'Block', 'bbox=', b.bbox.as_tuple())
            for l in b:
                print(' '*3, l.text.encode('UTF-8'), '(%0.2f, %0.2f, %0.2f, %0.2f)'% l.bbox.as_tuple())
                #assert l.char_fonts.comp_ratio < 1.0
                for i in range(len(l.text)):
                    print(l.text[i].encode('UTF-8'), '(%0.2f, %0.2f, %0.2f, %0.2f)'% l.char_bboxes[i].as_tuple(),\
                        l.char_fonts[i].name, l.char_fonts[i].size, l.char_fonts[i].color,)
                print()

你可以清楚地看到源代码是最短的，但仍然提供了所有必要的数据，包括字体颜色、字体大小、字体系列。

更重要的是，您可以直接在块 9 中获得单词，仅比字符高一级）。避免 space 检查逻辑，必须与 Pdfminer 一起使用。

单个单词的 PDFMiner 提取 - LTText LTTextBox

PDFMiner Extraction for Single Words - LTText LTTextBox

python

nlp

pdfminer