public StrongClassifier(WeakClassifier[] Classifiers)
{
this.Classifiers = Classifiers;
Func<double, WeakClassifier, double> sum = (acc, classifier) =>
acc + classifier.Alpha;
this.GlobalAlpha = this.Classifiers.Aggregate(0.0, sum);
}
/// <summary>
/// Construit un arbre de détection à partir d'un dossier
/// contenant des images de test valides et non valides.
/// </summary>
public static StrongClassifier Train(string TestsDir)
{
Console.WriteLine("Init trainer ...");
// Load default weights, integral images and compute features
// values foreach test image of the tests dirs.
// Training tests are resized to the window size.
Console.WriteLine("Load trainer tests ...");
var start = DateTime.Now;
var testsSet = LoadTestsSet(TestsDir);
var tests = testsSet.Item1;
var featuresValues = testsSet.Item2;
var nCachedFeatures = featuresValues.Where((feature) => feature.Values != null)
.Count();
Console.WriteLine("{0} tests loaded in {1} secs ({2}% cached)", tests.Length,
(DateTime.Now - start).TotalSeconds,
nCachedFeatures * 100 / featuresValues.Length);
Console.WriteLine("Ok. Let's start the neurons lazer powered firing machine ...");
WeakClassifier[] classifiers = new WeakClassifier[Config.LearnPass];
var iPass = 1;
while (iPass <= Config.LearnPass) {
start = DateTime.Now;
Console.WriteLine("{0}/{1} trainer pass ...", iPass, Config.LearnPass);
// Normalize the weights of the images to get
// a rational distribution
var weightsSum = tests.Aggregate(0.0, (acc, test) => test.Weight + acc);
var validWeight = 0.0;
for (var iTest = 0; iTest < tests.Length; iTest++) {
tests[iTest].Weight = tests[iTest].Weight / weightsSum;
if (tests[iTest].Valid)
validWeight += tests[iTest].Weight;
}
// Select the feature with the lowest error level for the new
// set of tests's weights
TestWeakClassifier best = new TestWeakClassifier(featuresValues[0], 0, 1, double.MaxValue);
//foreach (var feature in featuresValues) {
Parallel.ForEach(featuresValues, (feature) => {
var newClassifier = TestWeakClassifier.Train(tests, validWeight, feature);
if (best.Errors > newClassifier.Errors)
best = newClassifier;
});
//}
Console.WriteLine("New weak classifier selected in {0} secs (error score: {1})",
(DateTime.Now - start).TotalSeconds, best.Errors);
Console.WriteLine("X: {0} Y: {1} - Width : {2} Height : {3}",
best.Feature.Feature.Frame.TopLeft.X,
best.Feature.Feature.Frame.TopLeft.Y,
best.Feature.Feature.Frame.Width,
best.Feature.Feature.Frame.Height);
// Update the weights
var beta = best.Errors / (1.0 - best.Errors);
if (beta < 1e-8) // Unallow alpha to be > 18
beta = 1e-8;
// Reduce the weights of valid checks
foreach (var featureValue in best.Feature.Values) {
if (best.Check(featureValue.Value) == tests[featureValue.TestIndex].Valid)
tests[featureValue.TestIndex].Weight *= beta;
}
// Add the new weak classifier to the strong classifier
var alpha = Math.Log(1.0 / beta); // Weak classifier weight inside the strong classifier
classifiers[iPass - 1] = best.GetClassifier(alpha);
iPass += 1;
}
Console.WriteLine("Detector's brain ready to go");
return new StrongClassifier(classifiers);
}
