Source code for ferenda.pdfanalyze

# -*- coding: utf-8 -*-
from __future__ import unicode_literals
"""A bunch of helper functions for analyzing pdf files."""
import os
import logging
import json
from io import BytesIO, StringIO
from itertools import chain

from six import text_type as str

    import matplotlib
    import matplotlib.pyplot as plt
except ImportError:
    matplotlib = plt = None

    import PyPDF2
    from reportlab.pdfgen.canvas import Canvas
except ImportError:
    PyPDF2 = Canvas = None

from .pdfreader import Page
from ferenda import util
from ferenda.compat import Counter

[docs]class PDFAnalyzer(object): """Create a analyzer for the given pdf file. The primary purpose of an analyzer is to determine margins and other spatial metrics of a document, and identifiy common typographic styles for default text, title and headings. This is done by calling the :py:meth:`~ferenda.PDFAnalyzer.metrics` method. The analysis is done in several steps. The properties of all textboxes on each page is collected in several :py:class:`collections.Counter` objects. These counters are then statistically analyzed in a series of functions to yield these metrics. If different analyzis logic, or additional metrics, are desired, this class should be inherited and some methods/properties overridden. :param pdf: The pdf file to analyze. :type pdf: ferenda.PDFReader """ twopage = True """Whether or not the document is expected to have different margins depending on whether it's a even or odd page. """ frontmatter = 1 """The amount of frontmatter pages which might have differing typographic styles compared with the rest of the document. Affects style analysis, particularly how the title style is determined. """ style_significance_threshold = 0.005 """"The amount of use (as compared to the rest of the document that a style must have to be considered significant. """ header_significance_threshold = 0.002 """The maximum amount (expressed as part of the entire text amount) of text that can occur on the top of the page for it to be considered part of the header. """ # this is suitable for overriding, eg if you know that no pagenumbers occur in the footer footer_significance_threshold = 0.002 """The maximum amount (expressed as part of the entire text amount) of text that can occur on the bottom of the page for it to be considered part of the footer. """ frontmatter = 1 """The amount of pages to be considered frontmatter, which might have different typography, special title font etc.""" def __init__(self, pdf): # FIXME: in time, we'd like to make it possible to specify # multiple pdf files (either because a single logical document # is split into several files, or because the user wants to # analyze a bunch of docs in one go). self.pdf = pdf
[docs] def documents(self): """Attempts to distinguish different logical document (eg parts with differing pagesizes/margins/styles etc) within this PDF. You should override this method if you want to provide your own document segmentation logic. :returns: Tuples (startpage, pagecount) for the different identified documents :rtype: list """ return [(0, len(self.pdf))]
[docs] def metrics(self, metricspath=None, plotpath=None, startpage=0, pagecount=None, force=False): """Calculate and return the metrics for this analyzer. metrics is a set of named properties in the form of a dict. The keys of the dict can represent margins or other measurements of the document (left/right margins, header/footer etc) or font styles used in the document (eg. default, title, h1 -- h3). Style values are in turn dicts themselves, with the keys 'family' and 'size'. :param metricspath: The path of a JSON file used as cache for the calculated metrics :type metricspath: str :param plotpath: The path to write a PNG file with histograms for different values (for debugging). :type plotpath: str :param startpage: starting page for the analysis :type startpage: int :param startpage: number of pages to analyze (default: all available) :type startpage: int :param force: Perform analysis even if cached JSON metrics exists. :type force: bool :returns: calculated metrics :rtype: dict The default implementation will try to find out values for the following metrics: ================== =================================================== key description ================== =================================================== leftmargin position of left margin (for odd pages if twopage = True) rightmargin position of right margin (for odd pages if twopage = True) leftmargin_even position of left margin for even pages rightmargin_even position of right margin for right pages topmargin position of header zone bottommargin position of footer zone default style used for default text title style used for main document title (on front page) h1 style used for level 1 headings h2 style used for level 2 headings h3 style used for level 3 headings ================== =================================================== Subclasses might add (or remove) from the above. """ if (not force and metricspath and util.outfile_is_newer([self.pdf.filename], metricspath)): with open(metricspath) as fp: return json.load(fp) if pagecount is None: pagecount = len(self.pdf) - startpage hcounters = self.count_horizontal_margins(startpage, pagecount) vcounters = self.count_vertical_margins(startpage, pagecount) stylecounters = self.count_styles(startpage, pagecount) hmetrics = self.analyze_horizontal_margins(hcounters) vmetrics = self.analyze_vertical_margins(vcounters) stylemetrics = self.analyze_styles(stylecounters['frontmatter_styles'], stylecounters['rest_styles']) margincounters = dict(chain(hcounters.items(), vcounters.items())) allmetrics = dict(chain(hmetrics.items(), vmetrics.items(), stylemetrics.items())) if plotpath: self.plot(plotpath, margincounters, stylecounters, allmetrics) if metricspath: with open(metricspath, "w") as fp: json.dump(allmetrics, fp, indent=4) return allmetrics
[docs] def textboxes(self, startpage, pagecount): """Generate a stream of (pagenumber, textbox) tuples consisting of all pages/textboxes from startpage to pagecount. """ for page in self.pdf[startpage:startpage+pagecount]: for textbox in page: yield page.number, textbox
[docs] def count_horizontal_margins(self, startpage, pagecount): """Return a dict of Counter objects for all the horizontally oriented textbox properties (number of textboxes starting/ending at different positions). The set of counters is determined by setup_horizontal_counters. """ counters = self.setup_horizontal_counters() for pagenumber, textbox in self.textboxes(startpage, pagecount): self.count_horizontal_textbox(pagenumber, textbox, counters) for page in self.pdf[startpage:startpage+pagecount]: counters['pagewidth'][page.width] += 1 return counters
[docs] def setup_horizontal_counters(self): """Create initial set of horizontal counters.""" counters = {'leftmargin': Counter(), 'rightmargin': Counter(), 'pagewidth': Counter()} if self.twopage: counters['leftmargin_even'] = Counter() counters['rightmargin_even'] = Counter() return counters
[docs] def count_horizontal_textbox(self, pagenumber, textbox, counters): """Add a single textbox to the set of horizontal counters.""" if self.twopage and pagenumber % 2 == 0: counters['leftmargin_even'][textbox.left] += 1 counters['rightmargin_even'][textbox.right] += 1 else: counters['leftmargin'][textbox.left] += 1 counters['rightmargin'][textbox.right] += 1
[docs] def count_vertical_margins(self, startpage, pagecount): counters = self.setup_vertical_counters() for pagenumber, textbox in self.textboxes(startpage, pagecount): self.count_vertical_textbox(pagenumber, textbox, counters) for page in self.pdf[startpage:startpage+pagecount]: counters['pageheight'][page.height] += 1 return counters
[docs] def setup_vertical_counters(self): counters = {'topmargin': Counter(), 'bottommargin': Counter(), 'pageheight': Counter()} return counters
[docs] def count_vertical_textbox(self, pagenumber, textbox, counters): text = str(textbox).strip() counters['topmargin'][] += len(text) counters['bottommargin'][textbox.bottom] += len(text)
[docs] def count_styles(self, startpage, pagecount): counters = {'frontmatter_styles': Counter(), 'rest_styles': Counter()} for pagenumber, textbox in self.textboxes(startpage, pagecount): self.count_styles_textbox(pagenumber, textbox, counters) return counters
[docs] def count_styles_textbox(self, pagenumber, textbox, counters): text = str(textbox).strip() fonttuple = (, textbox.font.size) if pagenumber <= self.frontmatter: cid = "frontmatter_styles" else: cid = "rest_styles" counters[cid][fonttuple] += len(text)
[docs] def analyze_vertical_margins(self, vcounters): # now find probable header and footer zones. default algorithm: # max 0.2 % of text content can be in the header/footer zone. (on # a page of 2000 chars, only 4 can be in the footer) maxcount = self.header_significance_threshold * sum(vcounters['topmargin'].values()) charcount = 0 for i in range(max(vcounters['pageheight'])): charcount += vcounters['topmargin'].get(i, 0) if charcount > maxcount: header = i - 1 break charcount = 0 maxcount = self.footer_significance_threshold * sum(vcounters['topmargin'].values()) for i in range(max(vcounters['pageheight'])-1, -1, -1): charcount += vcounters['bottommargin'].get(i, 0) if charcount > maxcount: footer = i + 1 break return {'topmargin': header, 'bottommargin': footer}
# subclasses can (should) add metrics like parindent and secondcolumn
[docs] def analyze_horizontal_margins(self, vcounters): # find the probable left margin = the place where most textboxes start vmargins = {} pagewidth = vcounters['pagewidth'].most_common(1)[0][0] l = vcounters['leftmargin'] r = vcounters['rightmargin'] if l: vmargins['leftmargin'] = l.most_common(1)[0][0] assert vmargins['leftmargin'] < pagewidth / 2, "leftmargin shouldn't be on the right hand side of the page" if r: vmargins['rightmargin'] = r.most_common(1)[0][0] if self.twopage: le = vcounters['leftmargin_even'] re = vcounters['rightmargin_even'] if le: vmargins['leftmargin_even'] = le.most_common(1)[0][0] assert vmargins['leftmargin_even'] < pagewidth / 2, "leftmargin shouldn't be on the right hand side of the page" if re: vmargins['rightmargin_even'] = re.most_common(1)[0][0] vmargins['pagewidth'] = max(vcounters['pagewidth']) return vmargins
[docs] def fontsize_key(self, fonttuple): family, size = fonttuple if "Bold" in family: weight = 2 elif "Italic" in family: weight = 1 else: weight = 0 return (size, weight)
[docs] def fontdict(self, fonttuple): return {'family': fonttuple[0], 'size': fonttuple[1]}
[docs] def analyze_styles(self, frontmatter_styles, rest_styles): styledefs = {} all_styles = frontmatter_styles + rest_styles # default style: the one that's most common ds = all_styles.most_common(1)[0][0] styledefs['default'] = self.fontdict(ds) # title style: the largest style that exists on the frontpage style # objects are (should be) sortable. for style objects at the same # size, Bold > Italic > Regular if frontmatter_styles: ts = sorted(frontmatter_styles.keys(), key=self.fontsize_key, reverse=True)[0] styledefs['title'] = self.fontdict(ts) # h1 - h3: Take all styles larger than or equal to default, with # significant use (each > 0.5 % of all chars from page 2 onwards, # as the front page often uses nontypical styles), then order # styles by font size. significantuse = sum(rest_styles.values()) * 0.005 sortedstyles = sorted(rest_styles, key=self.fontsize_key, reverse=True) largestyles = [x for x in sortedstyles if (self.fontsize_key(x) > self.fontsize_key(ds) and rest_styles[x] > significantuse)] for style in ('h1', 'h2', 'h3'): if largestyles: # any left? styledefs[style] = self.fontdict(largestyles.pop(0)) return styledefs
[docs] def drawboxes(self, outfilename, gluefunc=None, startpage=0, pagecount=None, counters=None, metrics=None): """Create a copy of the parsed PDF file, but with the textboxes created by ``gluefunc`` clearly marked, and metrics shown on the page. .. note:: This requires PyPDF2 and reportlab, which aren't installed by default. Reportlab (3.*) only works on py27+ and py33+ """ if PyPDF2 is None or Canvas is None: raise ImportError("You need PyPDF2 and reportlab installed") styles = {} for k, v in metrics.items(): if isinstance(v, dict): styles[(v['family'], v['size'])] = k log = logging.getLogger("pdfanalyze") packet = None output = PyPDF2.PdfFileWriter() fp = open(self.pdf.filename, "rb") existing_pdf = PyPDF2.PdfFileReader(fp) pageidx = 0 tbidx = 0 sf = 2/3.0 # scaling factor -- mapping between units produced by # pdftohtml and units used by reportlab dirty = False for tb in self.pdf.textboxes(gluefunc, pageobjects=True): if isinstance(tb, Page): if dirty: new_pdf = PyPDF2.PdfFileReader(packet) # print("Merging a new page into %s" % id(existing_page)) # this SHOULD place the new page (that only # contains boxes and lines) on top of the existing # page. Only problem is that the image (from the # existing page) obscures those lines (contrary to # documentation) existing_page.mergePage(new_pdf.getPage(0)) output.addPage(existing_page) # alternate way: merge the existing page on top of # the new. This doesn't seem to work any better, # and creates issues with scaling. # # new_page = new_pdf.getPage(0) # new_page.mergePage(existing_page) # output.addPage(new_page) # print("Loading page %s" % pageidx) existing_page = existing_pdf.getPage(pageidx) pageidx += 1 mb = existing_page.mediaBox horizontal_scale = float(mb.getHeight()) / tb.height vertical_scale = float(mb.getWidth()) / tb.width log.debug("Loaded page %s - Scaling %s, %s" % (pageidx, horizontal_scale, vertical_scale)) packet = BytesIO() canvas = Canvas(packet, pagesize=(tb.height, tb.width), bottomup=False) # not sure how the vertical value 50 is derived... canvas.translate(0,50) canvas.scale(horizontal_scale, vertical_scale) canvas.setStrokeColorRGB(0.2,0.5,0.3) # now draw margins on the page for k, v in metrics.items(): if isinstance(v, int): if k in ('topmargin', 'bottommargin'): # horiz line canvas.line(0, v, tb.width, v) canvas.drawString(0, v, k) else: if ((k.endswith("_even") and (pageidx + 1) % 2 == 1) or (not k.endswith("_even") and (pageidx +1) % 2 == 0) or (not self.twopage)): # vert line canvas.line(v, 0, v, tb.height) canvas.drawString(v, tb.height-2, k) # for k, v in counters: # for each area in header, footer, leftmarg[_even], rightmarg[_even]: # select proper counter to draw in each area: # (headercounter->left gutter, # footercounter->right gutter, # leftmargcounter->header, # rightmargcounter->footer) # find the most frequent value in the selected counter, normalize agaist the space in the area # for each value of the counter draw single-width line at correct posiion tbidx = 0 else: tbidx += 1 dirty = True # for each textbox draw a square # add sequence number for textbox in lower right corner # if textbox style matches any in metrics, add the style name in upper lft corner canvas.rect(tb.left,, tb.width, tb.height) canvas.drawString(tb.left,, str(tbidx)) fonttuple = (, tb.font.size) if fonttuple in styles: canvas.drawString(tb.right, tb.bottom, styles[fonttuple]) new_pdf = PyPDF2.PdfFileReader(packet) # print("Merging final new page into %s" % id(existing_page)) existing_page.mergePage(new_pdf.getPage(0)) output.addPage(existing_page) outputStream = open(outfilename, "wb") output.write(outputStream) outputStream.close() log.debug("wrote %s" % outfilename) fp.close()
[docs] def plot(self, filename, margincounters, stylecounters, metrics): if matplotlib is None: raise ImportError("You need matplotlib installed") matplotlib.use('Agg') #'ggplot') # looks good but makes our histograms unreadable matplotlib.rcParams.update({'font.size': 8}) plt.figure(figsize=((len(margincounters))*2, 7)) # width, height in inches # if 6 counters: # +0,0--+ +0,1--+ +0,2--+ +0,3--+ # | LM | | LEM | | RM | | REM | # +-----+ +-----+ +-----+ +-----+ # +1,0--+ +1,1--+ +1,2 colspan=2+ # | TM | | BM | | Styles | # +-----+ +-----+ +-------------+ # # if 4 counters: # +0,0--+ +0,1--+ +0,2--+ # | LM | | RM | | TM | # +-----+ +-----+ +-----+ # +1,0--+ +1,1 colspan=2+ # | BM | | Styles | # +-----+ +-------------+ # disregard the pageheight/pagewidth counters pagewidth = max(margincounters['pagewidth']) del margincounters['pagewidth'] pageheight = max(margincounters['pageheight']) del margincounters['pageheight'] if len(margincounters) == 4: coords = ((0,0), (0,1), (0,2), (1,0), (1,1)) grid = (2,3) elif len(margincounters) == 6: coords = ((0,0), (0,1), (0,2), (0,3), (1,0), (1,1), (1,2)) grid = (2,4) else: # FIXME: make this dynamic raise ValueError("Can't layout other # of counters than 4 or 6") marginplots = [plt.subplot2grid(grid, pos) for pos in coords[:-1]] self.plot_margins(marginplots, margincounters, metrics, pagewidth, pageheight) styleplot = plt.subplot2grid(grid, coords[-1], colspan=2) self.plot_styles(styleplot, stylecounters, metrics) util.ensure_dir(filename) plt.savefig(filename, dpi=150) log = logging.getLogger("pdfanalyze") log.debug("wrote %s" % filename)
[docs] def plot_margins(self, subplots, margin_counters, metrics, pagewidth, pageheight): for (idx, counterkey) in enumerate(sorted(margin_counters.keys())): # print("Making plot for %s" % counterkey) # leftmargin_even => left # topmargin => top plot = subplots[idx] series = list(margin_counters[counterkey].elements()) size = pagewidth if "left" in counterkey or "right" in counterkey else pageheight # this is ridiculiously slow. Maybe ordinary bar charts # are faster? bins = plot.hist(series, bins=size, range=(0, size)) plot.set_title(counterkey) for k, v in metrics.items(): if counterkey == k: # FIXME: How annotate parindent, 2col etc ? label = "%s=%s" % (k, v) # leftmargin=102 # print(" plotting annotation %s" % label) plot.annotate(label, xy=(v, 100), xytext=(v*0.5, 100), arrowprops={'arrowstyle': '->'})
[docs] def plot_styles(self, plot, stylecounters, metrics): # do a additive vhist. FIXME: label those styles identified in # metrics allstyles = Counter(dict(chain(*[x.items() for x in stylecounters.values()]))) stylenames = [style[0].replace("TimesNewRomanPS", "Times")+"@"+str(style[1]) for style, count in allstyles.most_common()] stylecounts = [count for style, count in allstyles.most_common()] plt.yticks(range(len(stylenames)), stylenames) plot.barh(range(len(stylenames)), stylecounts, log=True) plot.set_title("Font usage", fontdict={'fontsize': 8})