fv/lib/find_text.js

FormVision is a node.js library for extracting data from scanned forms

// Generated by CoffeeScript 2.3.1
var boundingBox, boxDistanceVector, cloneUsingRegion, detectCandidates, detectLineMask, distanceVector, dv, findWords, intersectBox, isSameBlock, length, mergeRegions;

dv = require('dv');

({length, distanceVector, boxDistanceVector, intersectBox, boundingBox} = require('./math'));

// Compiles a mask with lines that have a certain length.
detectLineMask = function(image, minLineLength) {
  var k, len, line, lineMask, longLines;
  lineMask = new dv.Image(image.width, image.height, 8);
  longLines = image.toGray().lineSegments(0, 0, false).filter(function(line) {
    return length(distanceVector(line.p1, line.p2)) >= minLineLength;
  });
  for (k = 0, len = longLines.length; k < len; k++) {
    line = longLines[k];
    lineMask.drawLine(line.p1, line.p2, 7, 'set');
  }
  return lineMask;
};

mergeRegions = function(items, predicate) {
  var done, i, item, j, jj, k, l, len, len1, m, otherItem, ref, ref1, ref2, region, regions;
  // Initialize regions with unique indices.
  regions = (function() {
    var results = [];
    for (var k = 0, ref = items.length; 0 <= ref ? k < ref : k > ref; 0 <= ref ? k++ : k--){ results.push(k); }
    return results;
  }).apply(this);
  // Merge regions until predicate can no longer be applied.
  done = false;
  while (!done) {
    done = true;
// Merge regions (non-transitive).
    for (i = k = 0, len = items.length; k < len; i = ++k) {
      item = items[i];
      ref1 = items.slice(i + 1);
      for (j = l = 0, len1 = ref1.length; l < len1; j = ++l) {
        otherItem = ref1[j];
        jj = j + i + 1;
        if (regions[jj] !== regions[i] && predicate(item, otherItem)) {
          region = Math.min(regions[jj], regions[i]);
          regions[i] = regions[jj] = region;
          done = false;
        }
      }
    }
// Propagate merges (transitive).
    for (i = m = 0, ref2 = regions.length; (0 <= ref2 ? m <= ref2 : m >= ref2); i = 0 <= ref2 ? ++m : --m) {
      while (regions[regions[i]] !== regions[i]) {
        regions[i] = regions[regions[i]];
      }
    }
  }
  return regions;
};

isSameBlock = function(fontWidth, fontHeight) {
  return function(boxA, boxB) {
    var bottomA, bottomB, delta, sameLine;
    bottomA = boxA.y + boxA.height;
    bottomB = boxB.y + boxB.height;
    delta = boxDistanceVector(boxA, boxB);
    sameLine = Math.abs(bottomA - bottomB) < fontHeight / 2 && delta.x < fontWidth * 3;
    return sameLine || intersectBox(boxA, boxB);
  };
};

detectCandidates = function(binarizedImage, fontWidth = 20, fontHeight = 30) {
  var _, boxIndex, boxes, boxesByRegion, candidates, hasLetterSize, k, len, region, regions, smearHeight, smearWidth;
  hasLetterSize = function(box) {
    var ref;
    return fontWidth / 2 < box.width && (fontHeight / 2 < (ref = box.height) && ref < fontHeight * 6);
  };
  // Smear text a bit to extract letter boxes.
  smearWidth = (1 * fontWidth) + fontWidth % 2;
  smearHeight = (0.25 * fontHeight) + fontHeight % 2;
  boxes = binarizedImage.dilate(smearWidth, smearHeight).connectedComponents(8).filter(hasLetterSize);
  // Merge letter boxes to text regions.
  regions = mergeRegions(boxes, isSameBlock(fontWidth, fontHeight));
  boxesByRegion = {};
  for (boxIndex = k = 0, len = regions.length; k < len; boxIndex = ++k) {
    region = regions[boxIndex];
    if (boxesByRegion[region] == null) {
      boxesByRegion[region] = [];
    }
    boxesByRegion[region].push(boxes[boxIndex]);
  }
  candidates = (function() {
    var results;
    results = [];
    for (_ in boxesByRegion) {
      boxes = boxesByRegion[_];
      results.push(boxes);
    }
    return results;
  })();
  return candidates;
};

// Clone area of an image from boxes
cloneUsingRegion = function(image, boxes) {
  var box, cloneBox, cloneImage, k, len;
  cloneBox = boundingBox(boxes);
  cloneImage = new dv.Image(cloneBox.width, cloneBox.height, image.depth);
  cloneImage.clearBox({
    x: 0,
    y: 0,
    width: cloneBox.width,
    height: cloneBox.height
  });
  for (k = 0, len = boxes.length; k < len; k++) {
    box = boxes[k];
    cloneImage.drawImage(image.crop(box.x, box.y, box.width + 25, box.height), {
      x: box.x - cloneBox.x,
      y: box.y - cloneBox.y,
      width: box.width + 25,
      height: box.height
    });
  }
  return [cloneImage, cloneBox];
};

findWords = function(candidates, image, tesseract) {
  var candidateBoxes, cloneBox, cloneImage, k, l, len, len1, localWords, word, words;
  words = [];
  for (k = 0, len = candidates.length; k < len; k++) {
    candidateBoxes = candidates[k];
    // Crop and recognize.
    [cloneImage, cloneBox] = cloneUsingRegion(image, candidateBoxes);
    tesseract.image = cloneImage;
    tesseract.pageSegMode = cloneBox.height < 60 ? 'single_line' : 'single_block';
    localWords = tesseract.findWords();
    for (l = 0, len1 = localWords.length; l < len1; l++) {
      word = localWords[l];
      // Transform back.
      word.box.x += cloneBox.x;
      word.box.y += cloneBox.y;
      // Store candidate.
      word.candidate = candidateBoxes.slice(0);
    }
    words = words.concat(localWords);
  }
  // Filter words with tiny boxes.
  words = words.filter(function(word) {
    return word.box.width > 5 && word.box.height > 5;
  });
  return words;
};

// Use given *Tesseract* instance to find all text grouped as words along with
// confidence and boxes.
module.exports.findText = function(image, tesseract) {
  var candidates, clearedImage, k, len, lineMask, textImage, word, words;
  clearedImage = new dv.Image(image);
  // Remove long lines.
  lineMask = detectLineMask(image, 45);
  textImage = image.toGray().add(lineMask.toGray());
  // Find words using a simple Otsu thresholding.
  tesseract.image = textImage;
  candidates = detectCandidates(textImage.otsuAdaptiveThreshold(128, 128, 0, 0, 0).image);
  words = findWords(candidates, image, tesseract);
// Remove words from image, but safeguard against removing 'noise' that may be a checkmark.
  for (k = 0, len = words.length; k < len; k++) {
    word = words[k];
    if (word.text.length >= 6 || (word.text.length >= 3 && word.confidence >= 30)) {
      clearedImage.clearBox(word.box);
    }
  }
  return [words, clearedImage];
};