nfv/lib/find_text.js

[Updated to use ndv package instead dv] FormVision is a node.js library for extracting data from scanned forms

// Generated by CoffeeScript 1.12.7
var boundingBox, boxDistanceVector, cloneUsingRegion, detectCandidates, detectLineMask, distanceVector, dv, findWords, intersectBox, isSameBlock, length, mergeRegions, ref;

dv = require('ndv');

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

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, n, otherItem, ref1, ref2, ref3, region, regions, results;
  regions = (function() {
    results = [];
    for (var k = 0, ref1 = items.length; 0 <= ref1 ? k < ref1 : k > ref1; 0 <= ref1 ? k++ : k--){ results.push(k); }
    return results;
  }).apply(this);
  done = false;
  while (!done) {
    done = true;
    for (i = l = 0, len = items.length; l < len; i = ++l) {
      item = items[i];
      ref2 = items.slice(i + 1);
      for (j = m = 0, len1 = ref2.length; m < len1; j = ++m) {
        otherItem = ref2[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;
        }
      }
    }
    for (i = n = 0, ref3 = regions.length; 0 <= ref3 ? n <= ref3 : n >= ref3; i = 0 <= ref3 ? ++n : --n) {
      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, fontHeight) {
  var _, boxIndex, boxes, boxesByRegion, candidates, hasLetterSize, k, len, region, regions, smearHeight, smearWidth;
  if (fontWidth == null) {
    fontWidth = 20;
  }
  if (fontHeight == null) {
    fontHeight = 30;
  }
  hasLetterSize = function(box) {
    var ref1;
    return fontWidth / 2 < box.width && (fontHeight / 2 < (ref1 = box.height) && ref1 < fontHeight * 6);
  };
  smearWidth = (1 * fontWidth) + fontWidth % 2;
  smearHeight = (0.25 * fontHeight) + fontHeight % 2;
  boxes = binarizedImage.dilate(smearWidth, smearHeight).connectedComponents(8).filter(hasLetterSize);
  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;
};

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, ref1, word, words;
  words = [];
  for (k = 0, len = candidates.length; k < len; k++) {
    candidateBoxes = candidates[k];
    ref1 = cloneUsingRegion(image, candidateBoxes), cloneImage = ref1[0], cloneBox = ref1[1];
    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];
      word.box.x += cloneBox.x;
      word.box.y += cloneBox.y;
      word.candidate = candidateBoxes.slice(0);
    }
    words = words.concat(localWords);
  }
  words = words.filter(function(word) {
    return word.box.width > 5 && word.box.height > 5;
  });
  return words;
};

module.exports.findText = function(image, tesseract) {
  var candidates, clearedImage, k, len, lineMask, textImage, word, words;
  clearedImage = new dv.Image(image);
  lineMask = detectLineMask(image, 45);
  textImage = image.toGray().add(lineMask.toGray());
  tesseract.image = textImage;
  candidates = detectCandidates(textImage.otsuAdaptiveThreshold(128, 128, 0, 0, 0).image);
  words = findWords(candidates, image, tesseract);
  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];
};