tesseract/training/text2image.cpp

/**********************************************************************
 * File:        text2image.cpp
 * Description: Program to generate OCR training pages. Given a text file it
 *              outputs an image with a given font and degradation.
 *
 *              Note that since the results depend on the fonts available on
 *              your system, running the code on a different machine, or
 *              different OS, or even at a different time on the same machine,
 *              may produce different fonts even if --font is given explicitly.
 *              To see names of available fonts, use --list_available_fonts with
 *              the appropriate --fonts_dir path.
 *              Specifying --use_only_legacy_fonts will restrict the available
 *              fonts to those listed in legacy_fonts.h
 *
 * Authors:     Ranjith Unnikrishnan, Ray Smith
 * Created:     Tue Nov 19 2013
 *
 * (C) Copyright 2013, 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 <stdlib.h>
#include <string.h>
#include <algorithm>
#include <iostream>
#include <map>
#include <string>
#include <utility>
#include <vector>

#include "allheaders.h"  // from leptonica
#include "boxchar.h"
#include "commandlineflags.h"
#include "degradeimage.h"
#include "errcode.h"
#include "fileio.h"
#include "helpers.h"
#include "normstrngs.h"
#include "stringrenderer.h"
#include "tlog.h"
#include "unicharset.h"
#include "util.h"

#ifdef USE_STD_NAMESPACE
using std::make_pair;
using std::map;
using std::pair;
#endif

// A number with which to initialize the random number generator.
const int kRandomSeed = 0x18273645;

// The text input file.
STRING_PARAM_FLAG(text, "", "File name of text input to process");

// The text output file.
STRING_PARAM_FLAG(outputbase, "", "Basename for output image/box file");

// Degrade the rendered image to mimic scanner quality.
BOOL_PARAM_FLAG(degrade_image, true,
                "Degrade rendered image with speckle noise, dilation/erosion "
                "and rotation");

// Rotate the rendered image to have more realistic glyph borders
BOOL_PARAM_FLAG(rotate_image, true, "Rotate the image in a random way.");

// Degradation to apply to the image.
INT_PARAM_FLAG(exposure, 0, "Exposure level in photocopier");

// Output image resolution.
INT_PARAM_FLAG(resolution, 300, "Pixels per inch");

// Width of output image (in pixels).
INT_PARAM_FLAG(xsize, 3600, "Width of output image");

// Max height of output image (in pixels).
INT_PARAM_FLAG(ysize, 4800, "Height of output image");

// Margin around text (in pixels).
INT_PARAM_FLAG(margin, 100, "Margin round edges of image");

// Size of text (in points).
INT_PARAM_FLAG(ptsize, 12, "Size of printed text");

// Inter-character space (in ems).
DOUBLE_PARAM_FLAG(char_spacing, 0, "Inter-character space in ems");

// Sets the probability (value in [0, 1]) of starting to render a word with an
// underline. Words are assumed to be space-delimited.
DOUBLE_PARAM_FLAG(underline_start_prob, 0,
                  "Fraction of words to underline (value in [0,1])");
// Set the probability (value in [0, 1]) of continuing a started underline to
// the next word.
DOUBLE_PARAM_FLAG(underline_continuation_prob, 0,
                  "Fraction of words to underline (value in [0,1])");

// Inter-line space (in pixels).
INT_PARAM_FLAG(leading, 12, "Inter-line space (in pixels)");

// Layout and glyph orientation on rendering.
STRING_PARAM_FLAG(writing_mode, "horizontal",
                  "Specify one of the following writing"
                  " modes.\n"
                  "'horizontal' : Render regular horizontal text. (default)\n"
                  "'vertical' : Render vertical text. Glyph orientation is"
                  " selected by Pango.\n"
                  "'vertical-upright' : Render vertical text. Glyph "
                  " orientation is set to be upright.");

INT_PARAM_FLAG(box_padding, 0, "Padding around produced bounding boxes");

BOOL_PARAM_FLAG(strip_unrenderable_words, true,
                "Remove unrenderable words from source text");

// Font name.
STRING_PARAM_FLAG(font, "Arial", "Font description name to use");

BOOL_PARAM_FLAG(ligatures, false,
                "Rebuild and render ligatures");

BOOL_PARAM_FLAG(find_fonts, false,
                "Search for all fonts that can render the text");
BOOL_PARAM_FLAG(render_per_font, true,
                "If find_fonts==true, render each font to its own image. "
                "Image filenames are of the form output_name.font_name.tif");
DOUBLE_PARAM_FLAG(min_coverage, 1.0,
                  "If find_fonts==true, the minimum coverage the font has of "
                  "the characters in the text file to include it, between "
                  "0 and 1.");

BOOL_PARAM_FLAG(list_available_fonts, false, "List available fonts and quit.");

BOOL_PARAM_FLAG(render_ngrams, false, "Put each space-separated entity from the"
                " input file into one bounding box. The ngrams in the input"
                " file will be randomly permuted before rendering (so that"
                " there is sufficient variety of characters on each line).");

BOOL_PARAM_FLAG(output_word_boxes, false,
                "Output word bounding boxes instead of character boxes. "
                "This is used for Cube training, and implied by "
                "--render_ngrams.");

STRING_PARAM_FLAG(unicharset_file, "",
                  "File with characters in the unicharset. If --render_ngrams"
                  " is true and --unicharset_file is specified, ngrams with"
                  " characters that are not in unicharset will be omitted");

BOOL_PARAM_FLAG(bidirectional_rotation, false,
                "Rotate the generated characters both ways.");

BOOL_PARAM_FLAG(only_extract_font_properties, false,
                "Assumes that the input file contains a list of ngrams. Renders"
                " each ngram, extracts spacing properties and records them in"
                " output_base/[font_name].fontinfo file.");

// Use these flags to output zero-padded, square individual character images
BOOL_PARAM_FLAG(output_individual_glyph_images, false,
                "If true also outputs individual character images");
INT_PARAM_FLAG(glyph_resized_size, 0,
               "Each glyph is square with this side length in pixels");
INT_PARAM_FLAG(glyph_num_border_pixels_to_pad, 0,
               "Final_size=glyph_resized_size+2*glyph_num_border_pixels_to_pad");

namespace tesseract {

struct SpacingProperties {
  SpacingProperties() : x_gap_before(0), x_gap_after(0) {}
  SpacingProperties(int b, int a) : x_gap_before(b), x_gap_after(a) {}
  // These values are obtained from FT_Glyph_Metrics struct
  // used by the FreeType font engine.
  int x_gap_before;  // horizontal x bearing
  int x_gap_after;   // horizontal advance - x_gap_before - width
  map<string, int> kerned_x_gaps;
};

static bool IsWhitespaceBox(const BoxChar* boxchar) {
  return (boxchar->box() == NULL ||
          SpanUTF8Whitespace(boxchar->ch().c_str()));
}

static string StringReplace(const string& in,
                            const string& oldsub, const string& newsub) {
  string out;
  int start_pos = 0;
  do {
    int pos = in.find(oldsub, start_pos);
    if (pos == string::npos) break;
    out.append(in.data() + start_pos, pos - start_pos);
    out.append(newsub.data(), newsub.length());
    start_pos = pos + oldsub.length();
  } while (true);
  out.append(in.data() + start_pos, in.length() - start_pos);
  return out;
}

// Assumes that each word (whitespace-separated entity) in text is a bigram.
// Renders the bigrams and calls FontInfo::GetSpacingProperties() to
// obtain spacing information. Produces the output .fontinfo file with a line
// per unichar of the form:
// unichar space_before space_after kerned1 kerned_space1 kerned2 ...
// Fox example, if unichar "A" has spacing of 0 pixels before and -1 pixels
// after, is kerned with "V" resulting in spacing of "AV" to be -7 and kerned
// with "T", such that "AT" has spacing of -5, the entry/line for unichar "A"
// in .fontinfo file will be:
// A 0 -1 T -5 V -7
void ExtractFontProperties(const string &utf8_text,
                           StringRenderer *render,
                           const string &output_base) {
  map<string, SpacingProperties> spacing_map;
  map<string, SpacingProperties>::iterator spacing_map_it0;
  map<string, SpacingProperties>::iterator spacing_map_it1;
  int x_bearing, x_advance;
  int len = utf8_text.length();
  int offset = 0;
  const char* text = utf8_text.c_str();
  while (offset < len) {
    offset += render->RenderToImage(text + offset, strlen(text + offset), NULL);
    const vector<BoxChar*> &boxes = render->GetBoxes();

    // If the page break split a bigram, correct the offset so we try the bigram
    // on the next iteration.
    if (boxes.size() > 2 && !IsWhitespaceBox(boxes[boxes.size() - 1]) &&
        IsWhitespaceBox(boxes[boxes.size() - 2])) {
      if (boxes.size() > 3) {
        tprintf("WARNING: Adjusting to bad page break after '%s%s'\n",
                boxes[boxes.size() - 4]->ch().c_str(),
                boxes[boxes.size() - 3]->ch().c_str());
      }
      offset -= boxes[boxes.size() - 1]->ch().size();
    }

    for (int b = 0; b < boxes.size(); b += 2) {
      while (b < boxes.size() && IsWhitespaceBox(boxes[b])) ++b;
      if (b + 1 >= boxes.size()) break;
      const string &ch0 = boxes[b]->ch();
      // We encountered a ligature. This happens in at least two scenarios:
      // One is when the rendered bigram forms a grapheme cluster (eg. the
      // second character in the bigram is a combining vowel), in which case we
      // correctly output only one bounding box.
      // A second far less frequent case is when caused some fonts like 'DejaVu
      // Sans Ultra-Light' force Pango to render a ligatured character even if
      // the input consists of the separated characters.  NOTE(ranjith): As per
      // behdad@ this is not currently controllable at the level of the Pango
      // API.
      // The most frequent of all is a single character "word" made by the CJK
      // segmenter.
      // Safeguard against these cases here by just skipping the bigram.
      if (IsWhitespaceBox(boxes[b+1])) {
        continue;
      }
      int xgap = (boxes[b+1]->box()->x -
                  (boxes[b]->box()->x + boxes[b]->box()->w));
      spacing_map_it0 = spacing_map.find(ch0);
      int ok_count = 0;
      if (spacing_map_it0 == spacing_map.end() &&
          render->font().GetSpacingProperties(ch0, &x_bearing, &x_advance)) {
        spacing_map[ch0] = SpacingProperties(
            x_bearing, x_advance - x_bearing - boxes[b]->box()->w);
        spacing_map_it0 = spacing_map.find(ch0);
        ++ok_count;
      }
      const string &ch1 = boxes[b+1]->ch();
      tlog(3, "%s%s\n", ch0.c_str(), ch1.c_str());
      spacing_map_it1 = spacing_map.find(ch1);
      if (spacing_map_it1 == spacing_map.end() &&
          render->font().GetSpacingProperties(ch1, &x_bearing, &x_advance)) {
        spacing_map[ch1] = SpacingProperties(
            x_bearing, x_advance - x_bearing - boxes[b+1]->box()->w);
        spacing_map_it1 = spacing_map.find(ch1);
        ++ok_count;
      }
      if (ok_count == 2 && xgap != (spacing_map_it0->second.x_gap_after +
                                    spacing_map_it1->second.x_gap_before)) {
        spacing_map_it0->second.kerned_x_gaps[ch1] = xgap;
      }
    }
    render->ClearBoxes();
  }
  string output_string;
  const int kBufSize = 1024;
  char buf[kBufSize];
  snprintf(buf, kBufSize, "%d\n", static_cast<int>(spacing_map.size()));
  output_string.append(buf);
  map<string, SpacingProperties>::const_iterator spacing_map_it;
  for (spacing_map_it = spacing_map.begin();
       spacing_map_it != spacing_map.end(); ++spacing_map_it) {
    snprintf(buf, kBufSize,
             "%s %d %d %d", spacing_map_it->first.c_str(),
             spacing_map_it->second.x_gap_before,
             spacing_map_it->second.x_gap_after,
             static_cast<int>(spacing_map_it->second.kerned_x_gaps.size()));
    output_string.append(buf);
    map<string, int>::const_iterator kern_it;
    for (kern_it = spacing_map_it->second.kerned_x_gaps.begin();
         kern_it != spacing_map_it->second.kerned_x_gaps.end(); ++kern_it) {
      snprintf(buf, kBufSize,
               " %s %d", kern_it->first.c_str(), kern_it->second);
      output_string.append(buf);
    }
    output_string.append("\n");
  }
  File::WriteStringToFileOrDie(output_string, output_base + ".fontinfo");
}

bool MakeIndividualGlyphs(Pix* pix,
                          const vector<BoxChar*>& vbox,
                          const int input_tiff_page) {
  // If checks fail, return false without exiting text2image
  if (!pix) {
    tprintf("ERROR: MakeIndividualGlyphs(): Input Pix* is NULL\n");
    return false;
  } else if (FLAGS_glyph_resized_size <= 0) {
    tprintf("ERROR: --glyph_resized_size must be positive\n");
    return false;
  } else if (FLAGS_glyph_num_border_pixels_to_pad < 0) {
    tprintf("ERROR: --glyph_num_border_pixels_to_pad must be 0 or positive\n");
    return false;
  }

  const int n_boxes = vbox.size();
  int n_boxes_saved = 0;
  int current_tiff_page = 0;
  int y_previous = 0;
  static int glyph_count = 0;
  for (int i = 0; i < n_boxes; i++) {
    // Get one bounding box
    Box* b = vbox[i]->mutable_box();
    if (!b) continue;
    const int x = b->x;
    const int y = b->y;
    const int w = b->w;
    const int h = b->h;
    // Check present tiff page (for multipage tiff)
    if (y < y_previous-pixGetHeight(pix)/10) {
      tprintf("ERROR: Wrap-around encountered, at i=%d\n", i);
      current_tiff_page++;
    }
    if (current_tiff_page < input_tiff_page) continue;
    else if (current_tiff_page > input_tiff_page) break;
    // Check box validity
    if (x < 0 || y < 0 ||
        (x+w-1) >= pixGetWidth(pix) ||
        (y+h-1) >= pixGetHeight(pix)) {
      tprintf("ERROR: MakeIndividualGlyphs(): Index out of range, at i=%d"
              " (x=%d, y=%d, w=%d, h=%d\n)", i, x, y, w, h);
      continue;
    } else if (w < FLAGS_glyph_num_border_pixels_to_pad &&
               h < FLAGS_glyph_num_border_pixels_to_pad) {
      tprintf("ERROR: Input image too small to be a character, at i=%d\n", i);
      continue;
    }
    // Crop the boxed character
    Pix* pix_glyph = pixClipRectangle(pix, b, NULL);
    if (!pix_glyph) {
      tprintf("ERROR: MakeIndividualGlyphs(): Failed to clip, at i=%d\n", i);
      continue;
    }
    // Resize to square
    Pix* pix_glyph_sq = pixScaleToSize(pix_glyph,
                                       FLAGS_glyph_resized_size,
                                       FLAGS_glyph_resized_size);
    if (!pix_glyph_sq) {
      tprintf("ERROR: MakeIndividualGlyphs(): Failed to resize, at i=%d\n", i);
      continue;
    }
    // Zero-pad
    Pix* pix_glyph_sq_pad = pixAddBorder(pix_glyph_sq,
                                         FLAGS_glyph_num_border_pixels_to_pad,
                                         0);
    if (!pix_glyph_sq_pad) {
      tprintf("ERROR: MakeIndividualGlyphs(): Failed to zero-pad, at i=%d\n",
              i);
      continue;
    }
    // Write out
    Pix* pix_glyph_sq_pad_8 = pixConvertTo8(pix_glyph_sq_pad, false);
    char filename[1024];
    snprintf(filename, 1024, "%s_%d.jpg", FLAGS_outputbase.c_str(),
             glyph_count++);
    if (pixWriteJpeg(filename, pix_glyph_sq_pad_8, 100, 0)) {
      tprintf("ERROR: MakeIndividualGlyphs(): Failed to write JPEG to %s,"
              " at i=%d\n", filename, i);
      continue;
    }

    pixDestroy(&pix_glyph);
    pixDestroy(&pix_glyph_sq);
    pixDestroy(&pix_glyph_sq_pad);
    pixDestroy(&pix_glyph_sq_pad_8);
    n_boxes_saved++;
    y_previous = y;
  }
  if (n_boxes_saved == 0) {
    return false;
  } else {
    tprintf("Total number of characters saved = %d\n", n_boxes_saved);
    return true;
  }
}
}  // namespace tesseract

using tesseract::DegradeImage;
using tesseract::ExtractFontProperties;
using tesseract::File;
using tesseract::FontUtils;
using tesseract::SpanUTF8NotWhitespace;
using tesseract::SpanUTF8Whitespace;
using tesseract::StringRenderer;

int main(int argc, char** argv) {
  tesseract::ParseCommandLineFlags(argv[0], &argc, &argv, true);

  if (FLAGS_list_available_fonts) {
    const vector<string>& all_fonts = FontUtils::ListAvailableFonts();
    for (int i = 0; i < all_fonts.size(); ++i) {
      printf("%3d: %s\n", i, all_fonts[i].c_str());
      ASSERT_HOST_MSG(FontUtils::IsAvailableFont(all_fonts[i].c_str()),
                      "Font %s is unrecognized.\n", all_fonts[i].c_str());
    }
    return EXIT_SUCCESS;
  }

  // Check validity of input flags.
  if (FLAGS_text.empty()) {
    tprintf("'--text' option is missing!\n");
    exit(1);
  }
  if (FLAGS_outputbase.empty()) {
    tprintf("'--outputbase' option is missing!\n");
    exit(1);
  }
  if (!FLAGS_unicharset_file.empty() && FLAGS_render_ngrams) {
    tprintf("Use '--unicharset_file' only if '--render_ngrams' is set.\n");
    exit(1);
  }

  if (!FLAGS_find_fonts && !FontUtils::IsAvailableFont(FLAGS_font.c_str())) {
    string pango_name;
    if (!FontUtils::IsAvailableFont(FLAGS_font.c_str(), &pango_name)) {
      tprintf("Could not find font named %s.\n", FLAGS_font.c_str());
      if (!pango_name.empty()) {
        tprintf("Pango suggested font %s.\n", pango_name.c_str());
      }
      tprintf("Please correct --font arg.\n");
      exit(1);
    }
  }

  if (FLAGS_render_ngrams)
    FLAGS_output_word_boxes = true;

  char font_desc_name[1024];
  snprintf(font_desc_name, 1024, "%s %d", FLAGS_font.c_str(),
           static_cast<int>(FLAGS_ptsize));
  StringRenderer render(font_desc_name, FLAGS_xsize, FLAGS_ysize);
  render.set_add_ligatures(FLAGS_ligatures);
  render.set_leading(FLAGS_leading);
  render.set_resolution(FLAGS_resolution);
  render.set_char_spacing(FLAGS_char_spacing * FLAGS_ptsize);
  render.set_h_margin(FLAGS_margin);
  render.set_v_margin(FLAGS_margin);
  render.set_output_word_boxes(FLAGS_output_word_boxes);
  render.set_box_padding(FLAGS_box_padding);
  render.set_strip_unrenderable_words(FLAGS_strip_unrenderable_words);
  render.set_underline_start_prob(FLAGS_underline_start_prob);
  render.set_underline_continuation_prob(FLAGS_underline_continuation_prob);

  // Set text rendering orientation and their forms.
  if (FLAGS_writing_mode == "horizontal") {
    // Render regular horizontal text (default).
    render.set_vertical_text(false);
    render.set_gravity_hint_strong(false);
    render.set_render_fullwidth_latin(false);
  } else if (FLAGS_writing_mode == "vertical") {
    // Render vertical text. Glyph orientation is selected by Pango.
    render.set_vertical_text(true);
    render.set_gravity_hint_strong(false);
    render.set_render_fullwidth_latin(false);
  } else if (FLAGS_writing_mode == "vertical-upright") {
    // Render vertical text. Glyph orientation is set to be upright.
    // Also Basic Latin characters are converted to their fullwidth forms
    // on rendering, since fullwidth Latin characters are well designed to fit
    // vertical text lines, while .box files store halfwidth Basic Latin
    // unichars.
    render.set_vertical_text(true);
    render.set_gravity_hint_strong(true);
    render.set_render_fullwidth_latin(true);
  } else {
    tprintf("Invalid writing mode: %s\n", FLAGS_writing_mode.c_str());
    exit(1);
  }

  string src_utf8;
  // This c_str is NOT redundant!
  if (!File::ReadFileToString(FLAGS_text.c_str(), &src_utf8)) {
    tprintf("Failed to read file: %s\n", FLAGS_text.c_str());
    exit(1);
  }

  // Remove the unicode mark if present.
  if (strncmp(src_utf8.c_str(), "\xef\xbb\xbf", 3) == 0) {
    src_utf8.erase(0, 3);
  }
  tlog(1, "Render string of size %d\n", src_utf8.length());

  if (FLAGS_render_ngrams || FLAGS_only_extract_font_properties) {
    // Try to preserve behavior of old text2image by expanding inter-word
    // spaces by a factor of 4.
    const string kSeparator = FLAGS_render_ngrams ? "    " : " ";
    // Also restrict the number of charactes per line to try and avoid
    // line-breaking in the middle of words like "-A", "R$" etc. which are
    // otherwise allowed by the standard unicode line-breaking rules.
    const int kCharsPerLine = (FLAGS_ptsize > 20) ? 50 : 100;
    string rand_utf8;
    UNICHARSET unicharset;
    if (FLAGS_render_ngrams && !FLAGS_unicharset_file.empty() &&
        !unicharset.load_from_file(FLAGS_unicharset_file.c_str())) {
      tprintf("Failed to load unicharset from file %s\n",
              FLAGS_unicharset_file.c_str());
      exit(1);
    }

    // If we are rendering ngrams that will be OCRed later, shuffle them so that
    // tesseract does not have difficulties finding correct baseline, word
    // spaces, etc.
    const char *str8 = src_utf8.c_str();
    int len = src_utf8.length();
    int step;
    vector<pair<int, int> > offsets;
    int offset = SpanUTF8Whitespace(str8);
    while (offset < len) {
      step = SpanUTF8NotWhitespace(str8 + offset);
      offsets.push_back(make_pair(offset, step));
      offset += step;
      offset += SpanUTF8Whitespace(str8 + offset);
    }
    if (FLAGS_render_ngrams)
      std::random_shuffle(offsets.begin(), offsets.end());

    for (int i = 0, line = 1; i < offsets.size(); ++i) {
      const char *curr_pos = str8 + offsets[i].first;
      int ngram_len = offsets[i].second;
      // Skip words that contain characters not in found in unicharset.
      if (!FLAGS_unicharset_file.empty() &&
          !unicharset.encodable_string(curr_pos, NULL)) {
        continue;
      }
      rand_utf8.append(curr_pos, ngram_len);
      if (rand_utf8.length() > line * kCharsPerLine) {
        rand_utf8.append(" \n");
        ++line;
        if (line & 0x1) rand_utf8.append(kSeparator);
      } else {
        rand_utf8.append(kSeparator);
      }
    }
    tlog(1, "Rendered ngram string of size %d\n", rand_utf8.length());
    src_utf8.swap(rand_utf8);
  }
  if (FLAGS_only_extract_font_properties) {
    tprintf("Extracting font properties only\n");
    ExtractFontProperties(src_utf8, &render, FLAGS_outputbase.c_str());
    tprintf("Done!\n");
    return 0;
  }

  int im = 0;
  vector<float> page_rotation;
  const char* to_render_utf8 = src_utf8.c_str();

  tesseract::TRand randomizer;
  randomizer.set_seed(kRandomSeed);
  vector<string> font_names;
  // We use a two pass mechanism to rotate images in both direction.
  // The first pass(0) will rotate the images in random directions and
  // the second pass(1) will mirror those rotations.
  int num_pass = FLAGS_bidirectional_rotation ? 2 : 1;
  for (int pass = 0; pass < num_pass; ++pass) {
    int page_num = 0;
    string font_used;
    for (int offset = 0; offset < strlen(to_render_utf8); ++im, ++page_num) {
      tlog(1, "Starting page %d\n", im);
      Pix* pix = NULL;
      if (FLAGS_find_fonts) {
        offset += render.RenderAllFontsToImage(FLAGS_min_coverage,
                                               to_render_utf8 + offset,
                                               strlen(to_render_utf8 + offset),
                                               &font_used, &pix);
      } else {
        offset += render.RenderToImage(to_render_utf8 + offset,
                                       strlen(to_render_utf8 + offset), &pix);
      }
      if (pix != NULL) {
        float rotation = 0;
        if (pass == 1) {
          // Pass 2, do mirror rotation.
          rotation = -1 * page_rotation[page_num];
        }
        if (FLAGS_degrade_image) {
          pix = DegradeImage(pix, FLAGS_exposure, &randomizer,
                             FLAGS_rotate_image ? &rotation : NULL);
        }
        render.RotatePageBoxes(rotation);

        if (pass == 0) {
          // Pass 1, rotate randomly and store the rotation..
          page_rotation.push_back(rotation);
        }

        Pix* gray_pix = pixConvertTo8(pix, false);
        pixDestroy(&pix);
        Pix* binary = pixThresholdToBinary(gray_pix, 128);
        pixDestroy(&gray_pix);
        char tiff_name[1024];
        if (FLAGS_find_fonts) {
          if (FLAGS_render_per_font) {
            string fontname_for_file = tesseract::StringReplace(
                font_used, " ", "_");
            snprintf(tiff_name, 1024, "%s.%s.tif", FLAGS_outputbase.c_str(),
                     fontname_for_file.c_str());
            pixWriteTiff(tiff_name, binary, IFF_TIFF_G4, "w");
            tprintf("Rendered page %d to file %s\n", im, tiff_name);
          } else {
            font_names.push_back(font_used);
          }
        } else {
          snprintf(tiff_name, 1024, "%s.tif", FLAGS_outputbase.c_str());
          pixWriteTiff(tiff_name, binary, IFF_TIFF_G4, im == 0 ? "w" : "a");
          tprintf("Rendered page %d to file %s\n", im, tiff_name);
        }
        // Make individual glyphs
        if (FLAGS_output_individual_glyph_images) {
          if (!MakeIndividualGlyphs(binary, render.GetBoxes(), im)) {
            tprintf("ERROR: Individual glyphs not saved\n");
          }
        }
        pixDestroy(&binary);
      }
      if (FLAGS_find_fonts && offset != 0) {
        // We just want a list of names, or some sample images so we don't need
        // to render more than the first page of the text.
        break;
      }
    }
  }
  if (!FLAGS_find_fonts) {
    string box_name = FLAGS_outputbase.c_str();
    box_name += ".box";
    render.WriteAllBoxes(box_name);
  } else if (!FLAGS_render_per_font && !font_names.empty()) {
    string filename = FLAGS_outputbase.c_str();
    filename += ".fontlist.txt";
    FILE* fp = fopen(filename.c_str(), "wb");
    if (fp == NULL) {
      tprintf("Failed to create output font list %s\n", filename.c_str());
    } else {
      for (int i = 0; i < font_names.size(); ++i) {
        fprintf(fp, "%s\n", font_names[i].c_str());
      }
      fclose(fp);
    }
  }

  return 0;
}