public IEnumerable <(string Prediction, float Score, TextBlock Block)> Classify(Page page, IWordExtractor wordExtractor, IPageSegmenter pageSegmenter)
{
var words = wordExtractor.GetWords(page.Letters);
var blocks = pageSegmenter.GetBlocks(words);
foreach (var block in blocks)
{
var letters = block.TextLines.SelectMany(li => li.Words).SelectMany(w => w.Letters);
var paths = FeatureHelper.GetPathsInside(block.BoundingBox, page.ExperimentalAccess.Paths);
var images = FeatureHelper.GetImagesInside(block.BoundingBox, page.GetImages());
var features = FeatureHelper.GetFeatures(page, block.BoundingBox, letters, paths, images);
BlockFeatures blockFeatures = new BlockFeatures()
{
CharsCount = features[0],
PctNumericChars = features[1],
PctAlphabeticalChars = features[2],
PctSymbolicChars = features[3],
PctBulletChars = features[4],
DeltaToHeight = features[5],
PathsCount = features[6],
PctBezierPaths = features[7],
PctHorPaths = features[8],
PctVertPaths = features[9],
PctOblPaths = features[10],
ImagesCount = features[11],
ImageAvgProportion = features[12]
};
var result = predEngine.Predict(blockFeatures);
yield return(FeatureHelper.Categories[(int)result.Prediction], result.Score.Max(), block);
}
}
static void Main(string[] args)
{
// 1. Convert pdf documents and their PAGE xml ground truth to csv files
//DataGenerator.GetCsv(TEST_RAW_DATA_FILEPATH, 0, TEST_DATA_FILENAME); // testing
//DataGenerator.GetCsv(TRAIN_RAW_DATA_FILEPATH, 0, TRAIN_DATA_FILENAME); // training
// 2. Create the model
//LightGbmModelBuilder.TrainModel(DataGenerator.GetDataPath(TRAIN_DATA_FILENAME), MODEL_NAME);
// 3. Evaluate the model
//LightGbmModelBuilder.Evaluate(MODEL_NAME, DataGenerator.GetDataPath(TEST_DATA_FILENAME));
// 4. Load the trained classifier
LightGbmBlockClassifier lightGbmBlockClassifier = new LightGbmBlockClassifier(LightGbmModelBuilder.GetModelPath(MODEL_NAME));
var test = lightGbmBlockClassifier.OutputSchema["label"].HasSlotNames();
NearestNeighbourWordExtractor nearestNeighbourWordExtractor = new NearestNeighbourWordExtractor();
RecursiveXYCut recursiveXYCut = new RecursiveXYCut();
using (var document = PdfDocument.Open("sample.pdf"))
{
var hasBookmarks = document.TryGetBookmarks(out Bookmarks bookmarks);
for (var i = 0; i < document.NumberOfPages; i++)
{
var page = document.GetPage(i + 1);
List <DocumentBookmarkNode> bookmarksNodes = bookmarks?.GetNodes()
.Where(b => b is DocumentBookmarkNode)
.Select(b => b as DocumentBookmarkNode)
.Cast <DocumentBookmarkNode>()
.Where(b => b.PageNumber == page.Number).ToList();
var avgPageFontHeight = page.Letters.Select(l => l.GlyphRectangle.Height).Average();
var words = nearestNeighbourWordExtractor.GetWords(page.Letters);
var blocks = recursiveXYCut.GetBlocks(words, page.Width / 3.0);
foreach (var block in blocks)
{
var paths = FeatureHelper.GetPathsInside(block.BoundingBox, page.ExperimentalAccess.Paths);
var images = FeatureHelper.GetImagesInside(block.BoundingBox, page.GetImages());
var pred = lightGbmBlockClassifier.Classify(block, paths, images, avgPageFontHeight, bookmarksNodes);
Console.WriteLine();
Console.WriteLine(pred.Prediction + " [" + pred.Score.ToString("0.0%") + "]");
Console.WriteLine(block.Text.Normalize(normalizationForm: System.Text.NormalizationForm.FormKC)); // remove ligatures
}
}
}
Console.ReadKey();
}
public IEnumerable <(string Prediction, float Score, TextBlock Block)> Classify(Page page, IWordExtractor wordExtractor,
IPageSegmenter pageSegmenter, Bookmarks bookmarks = null)
{
List <DocumentBookmarkNode> bookmarksNodes = bookmarks?.GetNodes()
.Where(b => b is DocumentBookmarkNode)
.Select(b => b as DocumentBookmarkNode)
.Cast <DocumentBookmarkNode>()
.Where(b => b.PageNumber == page.Number).ToList();
var avgPageFontHeight = page.Letters.Select(l => l.GlyphRectangle.Height).Average();
var words = wordExtractor.GetWords(page.Letters);
var blocks = pageSegmenter.GetBlocks(words);
foreach (var block in blocks)
{
var letters = block.TextLines.SelectMany(li => li.Words).SelectMany(w => w.Letters);
var paths = FeatureHelper.GetPathsInside(block.BoundingBox, page.ExperimentalAccess.Paths);
var images = FeatureHelper.GetImagesInside(block.BoundingBox, page.GetImages());
var features = FeatureHelper.GetFeatures(
block, paths,
images, avgPageFontHeight,
block.BoundingBox.Area,
bookmarksNodes);
BlockFeatures blockFeatures = new BlockFeatures()
{
BlockAspectRatio = features[0],
CharsCount = features[1],
WordsCount = features[2],
LinesCount = features[3],
PctNumericChars = features[4],
PctAlphabeticalChars = features[5],
PctSymbolicChars = features[6],
PctBulletChars = features[7],
DeltaToHeight = features[8],
PathsCount = features[9],
PctBezierPaths = features[10],
PctHorPaths = features[11],
PctVertPaths = features[12],
PctOblPaths = features[13],
ImagesCount = features[14],
ImageAvgProportion = features[15],
BestNormEditDistance = features[16],
};
var result = predEngine.Predict(blockFeatures);
yield return(FeatureHelper.Categories[(int)result.Prediction], result.Score.Max(), block);
}
}