githubmirror/tesseract

src/api/pdfrenderer.cpp
976 строк32 KB
Amit Dovev
pdfrenderer.cpp: Ignore non-text blocks
08 ноя 2022, 09:02
08 ноя 2022, 09:02c196456
///////////////////////////////////////////////////////////////////////
// File:        pdfrenderer.cpp
// Description: PDF rendering interface to inject into TessBaseAPI
//
// (C) Copyright 2011, Google Inc.
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
// http://www.apache.org/licenses/LICENSE-2.0
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
///////////////////////////////////////////////////////////////////////

// Include automatically generated configuration file if running autoconf.
#ifdef HAVE_CONFIG_H
#  include "config_auto.h"
#endif

#include "pdf_ttf.h"
#include "tprintf.h"

#include <allheaders.h>
#include <tesseract/baseapi.h>
#include <tesseract/publictypes.h> // for PTIsTextType()
#include <tesseract/renderer.h>
#include <cmath>
#include <cstring>
#include <fstream>   // for std::ifstream
#include <locale>    // for std::locale::classic
#include <memory>    // std::unique_ptr
#include <sstream>   // for std::stringstream
#include "helpers.h" // for Swap

/*

Design notes from Ken Sharp, with light editing.

We think one solution is a font with a single glyph (.notdef) and a
CIDToGIDMap which maps all the CIDs to 0. That map would then be
stored as a stream in the PDF file, and when flat compressed should
be pretty small. The font, of course, will be approximately the same
size as the one you currently use.

I'm working on such a font now, the CIDToGIDMap is trivial, you just
create a stream object which contains 128k bytes (2 bytes per possible
CID and your CIDs range from 0 to 65535) and where you currently have
"/CIDToGIDMap /Identity" you would have "/CIDToGIDMap <object> 0 R".

Note that if, in future, you were to use a different (ie not 2 byte)
CMap for character codes you could trivially extend the CIDToGIDMap.

The following is an explanation of how some of the font stuff works,
this may be too simple for you in which case please accept my
apologies, its hard to know how much knowledge someone has. You can
skip all this anyway, its just for information.

The font embedded in a PDF file is usually intended just to be
rendered, but extensions allow for at least some ability to locate (or
copy) text from a document. This isn't something which was an original
goal of the PDF format, but its been retro-fitted, presumably due to
popular demand.

To do this reliably the PDF file must contain a ToUnicode CMap, a
device for mapping character codes to Unicode code points. If one of
these is present, then this will be used to convert the character
codes into Unicode values. If its not present then the reader will
fall back through a series of heuristics to try and guess the
result. This is, as you would expect, prone to failure.

This doesn't concern you of course, since you always write a ToUnicode
CMap, so because you are writing the text in text rendering mode 3 it
would seem that you don't really need to worry about this, but in the
PDF spec you cannot have an isolated ToUnicode CMap, it has to be
attached to a font, so in order to get even copy/paste to work you
need to define a font.

This is what leads to problems, tools like pdfwrite assume that they
are going to be able to (or even have to) modify the font entries, so
they require that the font being embedded be valid, and to be honest
the font Tesseract embeds isn't valid (for this purpose).


To see why lets look at how text is specified in a PDF file:

(Test) Tj

Now that looks like text but actually it isn't. Each of those bytes is
a 'character code'. When it comes to rendering the text a complex
sequence of events takes place, which converts the character code into
'something' which the font understands. Its entirely possible via
character mappings to have that text render as 'Sftu'

For simple fonts (PostScript type 1), we use the character code as the
index into an Encoding array (256 elements), each element of which is
a glyph name, so this gives us a glyph name. We then consult the
CharStrings dictionary in the font, that's a complex object which
contains pairs of keys and values, you can use the key to retrieve a
given value. So we have a glyph name, we then use that as the key to
the dictionary and retrieve the associated value. For a type 1 font,
the value is a glyph program that describes how to draw the glyph.

For CIDFonts, its a little more complicated. Because CIDFonts can be
large, using a glyph name as the key is unreasonable (it would also
lead to unfeasibly large Encoding arrays), so instead we use a 'CID'
as the key. CIDs are just numbers.

But.... We don't use the character code as the CID. What we do is use
a CMap to convert the character code into a CID. We then use the CID
to key the CharStrings dictionary and proceed as before. So the 'CMap'
is the equivalent of the Encoding array, but its a more compact and
flexible representation.

Note that you have to use the CMap just to find out how many bytes
constitute a character code, and it can be variable. For example you
can say if the first byte is 0x00->0x7f then its just one byte, if its
0x80->0xf0 then its 2 bytes and if its 0xf0->0xff then its 3 bytes. I
have seen CMaps defining character codes up to 5 bytes wide.

Now that's fine for 'PostScript' CIDFonts, but its not sufficient for
TrueType CIDFonts. The thing is that TrueType fonts are accessed using
a Glyph ID (GID) (and the LOCA table) which may well not be anything
like the CID. So for this case PDF includes a CIDToGIDMap. That maps
the CIDs to GIDs, and we can then use the GID to get the glyph
description from the GLYF table of the font.

So for a TrueType CIDFont, character-code->CID->GID->glyf-program.

Looking at the PDF file I was supplied with we see that it contains
text like :

<0x0075> Tj

So we start by taking the character code (117) and look it up in the
CMap. Well you don't supply a CMap, you just use the Identity-H one
which is predefined. So character code 117 maps to CID 117. Then we
use the CIDToGIDMap, again you don't supply one, you just use the
predefined 'Identity' map. So CID 117 maps to GID 117. But the font we
were supplied with only contains 116 glyphs.

Now for Latin that's not a huge problem, you can just supply a bigger
font. But for more complex languages that *is* going to be more of a
problem. Either you need to supply a font which contains glyphs for
all the possible CID->GID mappings, or we need to think laterally.

Our solution using a TrueType CIDFont is to intervene at the
CIDToGIDMap stage and convert all the CIDs to GID 0. Then we have a
font with just one glyph, the .notdef glyph at GID 0. This is what I'm
looking into now.

It would also be possible to have a 'PostScript' (ie type 1 outlines)
CIDFont which contained 1 glyph, and a CMap which mapped all character
codes to CID 0. The effect would be the same.

Its possible (I haven't checked) that the PostScript CIDFont and
associated CMap would be smaller than the TrueType font and associated
CIDToGIDMap.

--- in a followup ---

OK there is a small problem there, if I use GID 0 then Acrobat gets
upset about it and complains it cannot extract the font. If I set the
CIDToGIDMap so that all the entries are 1 instead, it's happy. Totally
mad......

*/

namespace tesseract {

// If the font is 10 pts, nominal character width is 5 pts
static const int kCharWidth = 2;

// Used for memory allocation. A codepoint must take no more than this
// many bytes, when written in the PDF way. e.g. "<0063>" for the
// letter 'c'
static const int kMaxBytesPerCodepoint = 20;

/**********************************************************************
 * PDF Renderer interface implementation
 **********************************************************************/
TessPDFRenderer::TessPDFRenderer(const char *outputbase, const char *datadir, bool textonly)
    : TessResultRenderer(outputbase, "pdf"), datadir_(datadir) {
  obj_ = 0;
  textonly_ = textonly;
  offsets_.push_back(0);
}

void TessPDFRenderer::AppendPDFObjectDIY(size_t objectsize) {
  offsets_.push_back(objectsize + offsets_.back());
  obj_++;
}

void TessPDFRenderer::AppendPDFObject(const char *data) {
  AppendPDFObjectDIY(strlen(data));
  AppendString(data);
}

// Helper function to prevent us from accidentally writing
// scientific notation to an HOCR or PDF file. Besides, three
// decimal points are all you really need.
static double prec(double x) {
  double kPrecision = 1000.0;
  double a = round(x * kPrecision) / kPrecision;
  if (a == -0) {
    return 0;
  }
  return a;
}

static long dist2(int x1, int y1, int x2, int y2) {
  return (x2 - x1) * (x2 - x1) + (y2 - y1) * (y2 - y1);
}

// Viewers like evince can get really confused during copy-paste when
// the baseline wanders around. So I've decided to project every word
// onto the (straight) line baseline. All numbers are in the native
// PDF coordinate system, which has the origin in the bottom left and
// the unit is points, which is 1/72 inch. Tesseract reports baselines
// left-to-right no matter what the reading order is. We need the
// word baseline in reading order, so we do that conversion here. Returns
// the word's baseline origin and length.
static void GetWordBaseline(int writing_direction, int ppi, int height, int word_x1, int word_y1,
                            int word_x2, int word_y2, int line_x1, int line_y1, int line_x2,
                            int line_y2, double *x0, double *y0, double *length) {
  if (writing_direction == WRITING_DIRECTION_RIGHT_TO_LEFT) {
    std::swap(word_x1, word_x2);
    std::swap(word_y1, word_y2);
  }
  double word_length;
  double x, y;
  {
    int px = word_x1;
    int py = word_y1;
    double l2 = dist2(line_x1, line_y1, line_x2, line_y2);
    if (l2 == 0) {
      x = line_x1;
      y = line_y1;
    } else {
      double t = ((px - line_x2) * (line_x2 - line_x1) + (py - line_y2) * (line_y2 - line_y1)) / l2;
      x = line_x2 + t * (line_x2 - line_x1);
      y = line_y2 + t * (line_y2 - line_y1);
    }
    word_length = sqrt(static_cast<double>(dist2(word_x1, word_y1, word_x2, word_y2)));
    word_length = word_length * 72.0 / ppi;
    x = x * 72 / ppi;
    y = height - (y * 72.0 / ppi);
  }
  *x0 = x;
  *y0 = y;
  *length = word_length;
}

// Compute coefficients for an affine matrix describing the rotation
// of the text. If the text is right-to-left such as Arabic or Hebrew,
// we reflect over the Y-axis. This matrix will set the coordinate
// system for placing text in the PDF file.
//
//                           RTL
// [ x' ] = [ a b ][ x ] = [-1 0 ] [ cos sin ][ x ]
// [ y' ]   [ c d ][ y ]   [ 0 1 ] [-sin cos ][ y ]
static void AffineMatrix(int writing_direction, int line_x1, int line_y1, int line_x2, int line_y2,
                         double *a, double *b, double *c, double *d) {
  double theta =
      atan2(static_cast<double>(line_y1 - line_y2), static_cast<double>(line_x2 - line_x1));
  *a = cos(theta);
  *b = sin(theta);
  *c = -sin(theta);
  *d = cos(theta);
  switch (writing_direction) {
    case WRITING_DIRECTION_RIGHT_TO_LEFT:
      *a = -*a;
      *b = -*b;
      break;
    case WRITING_DIRECTION_TOP_TO_BOTTOM:
      // TODO(jbreiden) Consider using the vertical PDF writing mode.
      break;
    default:
      break;
  }
}

// There are some really awkward PDF viewers in the wild, such as
// 'Preview' which ships with the Mac. They do a better job with text
// selection and highlighting when given perfectly flat baseline
// instead of very slightly tilted. We clip small tilts to appease
// these viewers. I chose this threshold large enough to absorb noise,
// but small enough that lines probably won't cross each other if the
// whole page is tilted at almost exactly the clipping threshold.
static void ClipBaseline(int ppi, int x1, int y1, int x2, int y2, int *line_x1, int *line_y1,
                         int *line_x2, int *line_y2) {
  *line_x1 = x1;
  *line_y1 = y1;
  *line_x2 = x2;
  *line_y2 = y2;
  int rise = abs(y2 - y1) * 72;
  int run = abs(x2 - x1) * 72;
  if (rise < 2 * ppi && 2 * ppi < run) {
    *line_y1 = *line_y2 = (y1 + y2) / 2;
  }
}

static bool CodepointToUtf16be(int code, char utf16[kMaxBytesPerCodepoint]) {
  if ((code > 0xD7FF && code < 0xE000) || code > 0x10FFFF) {
    tprintf("Dropping invalid codepoint %d\n", code);
    return false;
  }
  if (code < 0x10000) {
    snprintf(utf16, kMaxBytesPerCodepoint, "%04X", code);
  } else {
    int a = code - 0x010000;
    int high_surrogate = (0x03FF & (a >> 10)) + 0xD800;
    int low_surrogate = (0x03FF & a) + 0xDC00;
    snprintf(utf16, kMaxBytesPerCodepoint, "%04X%04X", high_surrogate, low_surrogate);
  }
  return true;
}

char *TessPDFRenderer::GetPDFTextObjects(TessBaseAPI *api, double width, double height) {
  double ppi = api->GetSourceYResolution();

  // These initial conditions are all arbitrary and will be overwritten
  double old_x = 0.0, old_y = 0.0;
  int old_fontsize = 0;
  tesseract::WritingDirection old_writing_direction = WRITING_DIRECTION_LEFT_TO_RIGHT;
  bool new_block = true;
  int fontsize = 0;
  double a = 1;
  double b = 0;
  double c = 0;
  double d = 1;

  std::stringstream pdf_str;
  // Use "C" locale (needed for double values prec()).
  pdf_str.imbue(std::locale::classic());
  // Use 8 digits for double values.
  pdf_str.precision(8);

  // TODO(jbreiden) This marries the text and image together.
  // Slightly cleaner from an abstraction standpoint if this were to
  // live inside a separate text object.
  pdf_str << "q " << prec(width) << " 0 0 " << prec(height) << " 0 0 cm";
  if (!textonly_) {
    pdf_str << " /Im1 Do";
  }
  pdf_str << " Q\n";

  int line_x1 = 0;
  int line_y1 = 0;
  int line_x2 = 0;
  int line_y2 = 0;

  const std::unique_ptr</*non-const*/ ResultIterator> res_it(api->GetIterator());
  while (!res_it->Empty(RIL_BLOCK)) {
    if (res_it->IsAtBeginningOf(RIL_BLOCK)) {
      auto block_type = res_it->BlockType();
      if (!PTIsTextType(block_type)) {
        // ignore non-text blocks
        res_it->Next(RIL_BLOCK);
        continue;
      }
      pdf_str << "BT\n3 Tr"; // Begin text object, use invisible ink
      old_fontsize = 0;      // Every block will declare its fontsize
      new_block = true;      // Every block will declare its affine matrix
    }

    if (res_it->IsAtBeginningOf(RIL_TEXTLINE)) {
      int x1, y1, x2, y2;
      res_it->Baseline(RIL_TEXTLINE, &x1, &y1, &x2, &y2);
      ClipBaseline(ppi, x1, y1, x2, y2, &line_x1, &line_y1, &line_x2, &line_y2);
    }

    if (res_it->Empty(RIL_WORD)) {
      res_it->Next(RIL_WORD);
      continue;
    }

    // Writing direction changes at a per-word granularity
    tesseract::WritingDirection writing_direction;
    {
      tesseract::Orientation orientation;
      tesseract::TextlineOrder textline_order;
      float deskew_angle;
      res_it->Orientation(&orientation, &writing_direction, &textline_order, &deskew_angle);
      if (writing_direction != WRITING_DIRECTION_TOP_TO_BOTTOM) {
        switch (res_it->WordDirection()) {
          case DIR_LEFT_TO_RIGHT:
            writing_direction = WRITING_DIRECTION_LEFT_TO_RIGHT;
            break;
          case DIR_RIGHT_TO_LEFT:
            writing_direction = WRITING_DIRECTION_RIGHT_TO_LEFT;
            break;
          default:
            writing_direction = old_writing_direction;
        }
      }
    }

    // Where is word origin and how long is it?
    double x, y, word_length;
    {
      int word_x1, word_y1, word_x2, word_y2;
      res_it->Baseline(RIL_WORD, &word_x1, &word_y1, &word_x2, &word_y2);
      GetWordBaseline(writing_direction, ppi, height, word_x1, word_y1, word_x2, word_y2, line_x1,
                      line_y1, line_x2, line_y2, &x, &y, &word_length);
    }

    if (writing_direction != old_writing_direction || new_block) {
      AffineMatrix(writing_direction, line_x1, line_y1, line_x2, line_y2, &a, &b, &c, &d);
      pdf_str << " " << prec(a) // . This affine matrix
              << " " << prec(b) // . sets the coordinate
              << " " << prec(c) // . system for all
              << " " << prec(d) // . text that follows.
              << " " << prec(x) // .
              << " " << prec(y) // .
              << (" Tm ");      // Place cursor absolutely
      new_block = false;
    } else {
      double dx = x - old_x;
      double dy = y - old_y;
      pdf_str << " " << prec(dx * a + dy * b) << " " << prec(dx * c + dy * d)
              << (" Td "); // Relative moveto
    }
    old_x = x;
    old_y = y;
    old_writing_direction = writing_direction;

    // Adjust font size on a per word granularity. Pay attention to
    // fontsize, old_fontsize, and pdf_str. We've found that for
    // in Arabic, Tesseract will happily return a fontsize of zero,
    // so we make up a default number to protect ourselves.
    {
      bool bold, italic, underlined, monospace, serif, smallcaps;
      int font_id;
      res_it->WordFontAttributes(&bold, &italic, &underlined, &monospace, &serif, &smallcaps,
                                 &fontsize, &font_id);
      const int kDefaultFontsize = 8;
      if (fontsize <= 0) {
        fontsize = kDefaultFontsize;
      }
      if (fontsize != old_fontsize) {
        pdf_str << "/f-0-0 " << fontsize << " Tf ";
        old_fontsize = fontsize;
      }
    }

    bool last_word_in_line = res_it->IsAtFinalElement(RIL_TEXTLINE, RIL_WORD);
    bool last_word_in_block = res_it->IsAtFinalElement(RIL_BLOCK, RIL_WORD);
    std::string pdf_word;
    int pdf_word_len = 0;
    do {
      const std::unique_ptr<const char[]> grapheme(res_it->GetUTF8Text(RIL_SYMBOL));
      if (grapheme && grapheme[0] != '\0') {
        std::vector<char32> unicodes = UNICHAR::UTF8ToUTF32(grapheme.get());
        char utf16[kMaxBytesPerCodepoint];
        for (char32 code : unicodes) {
          if (CodepointToUtf16be(code, utf16)) {
            pdf_word += utf16;
            pdf_word_len++;
          }
        }
      }
      res_it->Next(RIL_SYMBOL);
    } while (!res_it->Empty(RIL_BLOCK) && !res_it->IsAtBeginningOf(RIL_WORD));
    if (res_it->IsAtBeginningOf(RIL_WORD)) {
      pdf_word += "0020";
      pdf_word_len++;
    }
    if (word_length > 0 && pdf_word_len > 0) {
      double h_stretch = kCharWidth * prec(100.0 * word_length / (fontsize * pdf_word_len));
      pdf_str << h_stretch << " Tz" // horizontal stretch
              << " [ <" << pdf_word // UTF-16BE representation
              << "> ] TJ";          // show the text
    }
    if (last_word_in_line) {
      pdf_str << " \n";
    }
    if (last_word_in_block) {
      pdf_str << "ET\n"; // end the text object
    }
  }
  const std::string &text = pdf_str.str();
  char *result = new char[text.length() + 1];
  strcpy(result, text.c_str());
  return result;
}

bool TessPDFRenderer::BeginDocumentHandler() {
  AppendPDFObject("%PDF-1.5\n%\xDE\xAD\xBE\xEB\n");

  // CATALOG
  AppendPDFObject(
      "1 0 obj\n"
      "<<\n"
      "  /Type /Catalog\n"
      "  /Pages 2 0 R\n"
      ">>\nendobj\n");

  // We are reserving object #2 for the /Pages
  // object, which I am going to create and write
  // at the end of the PDF file.
  AppendPDFObject("");

  // TYPE0 FONT
  AppendPDFObject(
      "3 0 obj\n"
      "<<\n"
      "  /BaseFont /GlyphLessFont\n"
      "  /DescendantFonts [ 4 0 R ]\n" // CIDFontType2 font
      "  /Encoding /Identity-H\n"
      "  /Subtype /Type0\n"
      "  /ToUnicode 6 0 R\n" // ToUnicode
      "  /Type /Font\n"
      ">>\n"
      "endobj\n");

  // CIDFONTTYPE2
  std::stringstream stream;
  // Use "C" locale (needed for int values larger than 999).
  stream.imbue(std::locale::classic());
  stream << "4 0 obj\n"
            "<<\n"
            "  /BaseFont /GlyphLessFont\n"
            "  /CIDToGIDMap 5 0 R\n" // CIDToGIDMap
            "  /CIDSystemInfo\n"
            "  <<\n"
            "     /Ordering (Identity)\n"
            "     /Registry (Adobe)\n"
            "     /Supplement 0\n"
            "  >>\n"
            "  /FontDescriptor 7 0 R\n" // Font descriptor
            "  /Subtype /CIDFontType2\n"
            "  /Type /Font\n"
            "  /DW "
         << (1000 / kCharWidth)
         << "\n"
            ">>\n"
            "endobj\n";
  AppendPDFObject(stream.str().c_str());

  // CIDTOGIDMAP
  const int kCIDToGIDMapSize = 2 * (1 << 16);
  const std::unique_ptr<unsigned char[]> cidtogidmap(new unsigned char[kCIDToGIDMapSize]);
  for (int i = 0; i < kCIDToGIDMapSize; i++) {
    cidtogidmap[i] = (i % 2) ? 1 : 0;
  }
  size_t len;
  unsigned char *comp = zlibCompress(cidtogidmap.get(), kCIDToGIDMapSize, &len);
  stream.str("");
  stream << "5 0 obj\n"
            "<<\n"
            "  /Length "
         << len
         << " /Filter /FlateDecode\n"
            ">>\n"
            "stream\n";
  AppendString(stream.str().c_str());
  long objsize = stream.str().size();
  AppendData(reinterpret_cast<char *>(comp), len);
  objsize += len;
  lept_free(comp);
  const char *endstream_endobj =
      "endstream\n"
      "endobj\n";
  AppendString(endstream_endobj);
  objsize += strlen(endstream_endobj);
  AppendPDFObjectDIY(objsize);

  const char stream2[] =
      "/CIDInit /ProcSet findresource begin\n"
      "12 dict begin\n"
      "begincmap\n"
      "/CIDSystemInfo\n"
      "<<\n"
      "  /Registry (Adobe)\n"
      "  /Ordering (UCS)\n"
      "  /Supplement 0\n"
      ">> def\n"
      "/CMapName /Adobe-Identify-UCS def\n"
      "/CMapType 2 def\n"
      "1 begincodespacerange\n"
      "<0000> <FFFF>\n"
      "endcodespacerange\n"
      "1 beginbfrange\n"
      "<0000> <FFFF> <0000>\n"
      "endbfrange\n"
      "endcmap\n"
      "CMapName currentdict /CMap defineresource pop\n"
      "end\n"
      "end\n";

  // TOUNICODE
  stream.str("");
  stream << "6 0 obj\n"
            "<< /Length "
         << (sizeof(stream2) - 1)
         << " >>\n"
            "stream\n"
         << stream2
         << "endstream\n"
            "endobj\n";
  AppendPDFObject(stream.str().c_str());

  // FONT DESCRIPTOR
  stream.str("");
  stream << "7 0 obj\n"
            "<<\n"
            "  /Ascent 1000\n"
            "  /CapHeight 1000\n"
            "  /Descent -1\n" // Spec says must be negative
            "  /Flags 5\n"    // FixedPitch + Symbolic
            "  /FontBBox  [ 0 0 "
         << (1000 / kCharWidth)
         << " 1000 ]\n"
            "  /FontFile2 8 0 R\n"
            "  /FontName /GlyphLessFont\n"
            "  /ItalicAngle 0\n"
            "  /StemV 80\n"
            "  /Type /FontDescriptor\n"
            ">>\n"
            "endobj\n";
  AppendPDFObject(stream.str().c_str());

  stream.str("");
  stream << datadir_.c_str() << "/pdf.ttf";
  const uint8_t *font;
  std::ifstream input(stream.str().c_str(), std::ios::in | std::ios::binary);
  std::vector<unsigned char> buffer(std::istreambuf_iterator<char>(input), {});
  auto size = buffer.size();
  if (size) {
    font = buffer.data();
  } else {
#if !defined(NDEBUG)
    tprintf("Cannot open file \"%s\"!\nUsing internal glyphless font.\n", stream.str().c_str());
#endif
    font = pdf_ttf;
    size = sizeof(pdf_ttf);
  }

  // FONTFILE2
  stream.str("");
  stream << "8 0 obj\n"
            "<<\n"
            "  /Length "
         << size
         << "\n"
            "  /Length1 "
         << size
         << "\n"
            ">>\n"
            "stream\n";
  AppendString(stream.str().c_str());
  objsize = stream.str().size();
  AppendData(reinterpret_cast<const char *>(font), size);
  objsize += size;
  AppendString(endstream_endobj);
  objsize += strlen(endstream_endobj);
  AppendPDFObjectDIY(objsize);
  return true;
}

bool TessPDFRenderer::imageToPDFObj(Pix *pix, const char *filename, long int objnum,
                                    char **pdf_object, long int *pdf_object_size,
                                    const int jpg_quality) {
  if (!pdf_object_size || !pdf_object) {
    return false;
  }
  *pdf_object = nullptr;
  *pdf_object_size = 0;
  if (!filename && !pix) {
    return false;
  }

  L_Compressed_Data *cid = nullptr;

  int sad = 0;
  if (pixGetInputFormat(pix) == IFF_PNG) {
    sad = pixGenerateCIData(pix, L_FLATE_ENCODE, 0, 0, &cid);
  }
  if (!cid) {
    sad = l_generateCIDataForPdf(filename, pix, jpg_quality, &cid);
  }

  if (sad || !cid) {
    l_CIDataDestroy(&cid);
    return false;
  }

  const char *group4 = "";
  const char *filter;
  switch (cid->type) {
    case L_FLATE_ENCODE:
      filter = "/FlateDecode";
      break;
    case L_JPEG_ENCODE:
      filter = "/DCTDecode";
      break;
    case L_G4_ENCODE:
      filter = "/CCITTFaxDecode";
      group4 = "    /K -1\n";
      break;
    case L_JP2K_ENCODE:
      filter = "/JPXDecode";
      break;
    default:
      l_CIDataDestroy(&cid);
      return false;
  }

  // Maybe someday we will accept RGBA but today is not that day.
  // It requires creating an /SMask for the alpha channel.
  // http://stackoverflow.com/questions/14220221
  std::stringstream colorspace;
  // Use "C" locale (needed for int values larger than 999).
  colorspace.imbue(std::locale::classic());
  if (cid->ncolors > 0) {
    colorspace << "  /ColorSpace [ /Indexed /DeviceRGB " << (cid->ncolors - 1) << " "
               << cid->cmapdatahex << " ]\n";
  } else {
    switch (cid->spp) {
      case 1:
        if (cid->bps == 1 && pixGetInputFormat(pix) == IFF_PNG) {
          colorspace.str(
              "  /ColorSpace /DeviceGray\n"
              "  /Decode [1 0]\n");
        } else {
          colorspace.str("  /ColorSpace /DeviceGray\n");
        }
        break;
      case 3:
        colorspace.str("  /ColorSpace /DeviceRGB\n");
        break;
      default:
        l_CIDataDestroy(&cid);
        return false;
    }
  }

  int predictor = (cid->predictor) ? 14 : 1;

  // IMAGE
  std::stringstream b1;
  // Use "C" locale (needed for int values larger than 999).
  b1.imbue(std::locale::classic());
  b1 << objnum
     << " 0 obj\n"
        "<<\n"
        "  /Length "
     << cid->nbytescomp
     << "\n"
        "  /Subtype /Image\n";

  std::stringstream b2;
  // Use "C" locale (needed for int values larger than 999).
  b2.imbue(std::locale::classic());
  b2 << "  /Width " << cid->w
     << "\n"
        "  /Height "
     << cid->h
     << "\n"
        "  /BitsPerComponent "
     << cid->bps
     << "\n"
        "  /Filter "
     << filter
     << "\n"
        "  /DecodeParms\n"
        "  <<\n"
        "    /Predictor "
     << predictor
     << "\n"
        "    /Colors "
     << cid->spp << "\n"
     << group4 << "    /Columns " << cid->w
     << "\n"
        "    /BitsPerComponent "
     << cid->bps
     << "\n"
        "  >>\n"
        ">>\n"
        "stream\n";

  const char *b3 =
      "endstream\n"
      "endobj\n";

  size_t b1_len = b1.str().size();
  size_t b2_len = b2.str().size();
  size_t b3_len = strlen(b3);
  size_t colorspace_len = colorspace.str().size();

  *pdf_object_size = b1_len + colorspace_len + b2_len + cid->nbytescomp + b3_len;
  *pdf_object = new char[*pdf_object_size];

  char *p = *pdf_object;
  memcpy(p, b1.str().c_str(), b1_len);
  p += b1_len;
  memcpy(p, colorspace.str().c_str(), colorspace_len);
  p += colorspace_len;
  memcpy(p, b2.str().c_str(), b2_len);
  p += b2_len;
  memcpy(p, cid->datacomp, cid->nbytescomp);
  p += cid->nbytescomp;
  memcpy(p, b3, b3_len);
  l_CIDataDestroy(&cid);
  return true;
}

bool TessPDFRenderer::AddImageHandler(TessBaseAPI *api) {
  Pix *pix = api->GetInputImage();
  const char *filename = api->GetInputName();
  int ppi = api->GetSourceYResolution();
  if (!pix || ppi <= 0) {
    return false;
  }
  double width = pixGetWidth(pix) * 72.0 / ppi;
  double height = pixGetHeight(pix) * 72.0 / ppi;

  std::stringstream xobject;
  // Use "C" locale (needed for int values larger than 999).
  xobject.imbue(std::locale::classic());
  if (!textonly_) {
    xobject << "/XObject << /Im1 " << (obj_ + 2) << " 0 R >>\n";
  }

  // PAGE
  std::stringstream stream;
  // Use "C" locale (needed for double values width and height).
  stream.imbue(std::locale::classic());
  stream.precision(2);
  stream << std::fixed << obj_
         << " 0 obj\n"
            "<<\n"
            "  /Type /Page\n"
            "  /Parent 2 0 R\n" // Pages object
            "  /MediaBox [0 0 "
         << width << " " << height
         << "]\n"
            "  /Contents "
         << (obj_ + 1)
         << " 0 R\n" // Contents object
            "  /Resources\n"
            "  <<\n"
            "    "
         << xobject.str() << // Image object
      "    /ProcSet [ /PDF /Text /ImageB /ImageI /ImageC ]\n"
      "    /Font << /f-0-0 3 0 R >>\n" // Type0 Font
      "  >>\n"
      ">>\n"
      "endobj\n";
  pages_.push_back(obj_);
  AppendPDFObject(stream.str().c_str());

  // CONTENTS
  const std::unique_ptr<char[]> pdftext(GetPDFTextObjects(api, width, height));
  const size_t pdftext_len = strlen(pdftext.get());
  size_t len;
  unsigned char *comp_pdftext =
      zlibCompress(reinterpret_cast<unsigned char *>(pdftext.get()), pdftext_len, &len);
  long comp_pdftext_len = len;
  stream.str("");
  stream << obj_
         << " 0 obj\n"
            "<<\n"
            "  /Length "
         << comp_pdftext_len
         << " /Filter /FlateDecode\n"
            ">>\n"
            "stream\n";
  AppendString(stream.str().c_str());
  long objsize = stream.str().size();
  AppendData(reinterpret_cast<char *>(comp_pdftext), comp_pdftext_len);
  objsize += comp_pdftext_len;
  lept_free(comp_pdftext);
  const char *b2 =
      "endstream\n"
      "endobj\n";
  AppendString(b2);
  objsize += strlen(b2);
  AppendPDFObjectDIY(objsize);

  if (!textonly_) {
    char *pdf_object = nullptr;
    int jpg_quality;
    api->GetIntVariable("jpg_quality", &jpg_quality);
    if (!imageToPDFObj(pix, filename, obj_, &pdf_object, &objsize, jpg_quality)) {
      return false;
    }
    AppendData(pdf_object, objsize);
    AppendPDFObjectDIY(objsize);
    delete[] pdf_object;
  }
  return true;
}

bool TessPDFRenderer::EndDocumentHandler() {
  // We reserved the /Pages object number early, so that the /Page
  // objects could refer to their parent. We finally have enough
  // information to go fill it in. Using lower level calls to manipulate
  // the offset record in two spots, because we are placing objects
  // out of order in the file.

  // PAGES
  const long int kPagesObjectNumber = 2;
  offsets_[kPagesObjectNumber] = offsets_.back(); // manipulation #1
  std::stringstream stream;
  // Use "C" locale (needed for int values larger than 999).
  stream.imbue(std::locale::classic());
  stream << kPagesObjectNumber << " 0 obj\n<<\n  /Type /Pages\n  /Kids [ ";
  AppendString(stream.str().c_str());
  size_t pages_objsize = stream.str().size();
  for (const auto &page : pages_) {
    stream.str("");
    stream << page << " 0 R ";
    AppendString(stream.str().c_str());
    pages_objsize += stream.str().size();
  }
  stream.str("");
  stream << "]\n  /Count " << pages_.size() << "\n>>\nendobj\n";
  AppendString(stream.str().c_str());
  pages_objsize += stream.str().size();
  offsets_.back() += pages_objsize; // manipulation #2

  // INFO
  std::string utf16_title = "FEFF"; // byte_order_marker
  std::vector<char32> unicodes = UNICHAR::UTF8ToUTF32(title());
  char utf16[kMaxBytesPerCodepoint];
  for (char32 code : unicodes) {
    if (CodepointToUtf16be(code, utf16)) {
      utf16_title += utf16;
    }
  }

  char *datestr = l_getFormattedDate();
  stream.str("");
  stream << obj_
         << " 0 obj\n"
            "<<\n"
            "  /Producer (Tesseract "
         << tesseract::TessBaseAPI::Version()
         << ")\n"
            "  /CreationDate (D:"
         << datestr
         << ")\n"
            "  /Title <"
         << utf16_title.c_str()
         << ">\n"
            ">>\n"
            "endobj\n";
  lept_free(datestr);
  AppendPDFObject(stream.str().c_str());
  stream.str("");
  stream << "xref\n0 " << obj_ << "\n0000000000 65535 f \n";
  AppendString(stream.str().c_str());
  for (int i = 1; i < obj_; i++) {
    stream.str("");
    stream.width(10);
    stream.fill('0');
    stream << offsets_[i] << " 00000 n \n";
    AppendString(stream.str().c_str());
  }
  stream.str("");
  stream << "trailer\n<<\n  /Size " << obj_
         << "\n"
            "  /Root 1 0 R\n" // catalog
            "  /Info "
         << (obj_ - 1)
         << " 0 R\n" // info
            ">>\nstartxref\n"
         << offsets_.back() << "\n%%EOF\n";
  AppendString(stream.str().c_str());
  return true;
}
} // namespace tesseract