public void Train(double[][,] images, List <WeakClassifier> weakClassifiers, bool[] results, int featureCount)
{
if (results.Count(r => r) == 0)
{
return;
}
if (results.Count(r => !r) == 0)
{
return;
}
var integralImages = IntegrateImages(images);
var negativeCount = results.ToList().Where(r => r == false).Count();
var positiveCount = results.ToList().Where(r => r == true).Count();
WeakClassifiers = new List <WeakClassifier>();
Betas = new List <double>();
Weights = new List <double>(); // Strong classifier Weights
InitializeWeights(results);
Console.WriteLine("Training started...");
WeakClassifier minFeature = null;
// Weak classifiers loop
for (int featureIndex = 0; featureIndex < featureCount; featureIndex++)
{
// Normalize the weights
var weightSum = ImagesWeights[featureIndex].Sum();
for (int i = 0; i < images.Length; i++)
{
ImagesWeights[featureIndex][i] /= weightSum;
}
// Train each weak classifier in parallel
weakClassifiers.AsParallel().ForAll(wc =>
{
wc.Train(ref integralImages, ref results, ImagesWeights[featureIndex]);
});
// Choosing weak classifier with lowest error
minFeature = weakClassifiers.OrderBy(wc => wc.Error).First();
// Saving weak classifier
WeakClassifiers.Add(minFeature);
weakClassifiers.Remove(minFeature);
// Calc Betas and Weights
Betas.Add(minFeature.Error / (1 - minFeature.Error));
Weights.Add(Math.Log(1.0 / Betas[featureIndex]));
// Weights update
ImagesWeights.Add(new double[integralImages.Length]);
for (int i = 0; i < integralImages.Length; i++)
{
var featureResult = minFeature.Detect(integralImages[i]);
if (featureResult == results[i])
{
ImagesWeights[featureIndex + 1][i] = ImagesWeights[featureIndex][i] * Betas[featureIndex];
}
else
{
ImagesWeights[featureIndex + 1][i] = ImagesWeights[featureIndex][i];
}
}
Console.WriteLine("Feature #{0}. Error: {1}", featureIndex + 1, minFeature.Error);
}
var scores = CalcScores(integralImages, results);
DetermineThreshold(ref scores);
Statistics(images, results);
Console.WriteLine("Training finished.");
}
public void AddWeakClassifier(double[][,] images, bool[] results, List <WeakClassifier> weakClassifiers)
{
if (results.Count(r => r) == 0)
{
return;
}
if (results.Count(r => !r) == 0)
{
return;
}
if (ImagesWeights == null)
{
ImagesWeights = new List <double[]>();
}
if (Weights == null)
{
Weights = new List <double>();
}
if (WeakClassifiers == null)
{
WeakClassifiers = new List <WeakClassifier>();
}
if (Betas == null)
{
Betas = new List <double>();
}
if (ImagesWeights.Count == 0)
{
InitializeWeights(results);
}
var integralImages = IntegrateImages(images);
var negativeCount = results.ToList().Where(r => r == false).Count();
var positiveCount = results.ToList().Where(r => r == true).Count();
WeakClassifier minFeature = null;
// Normalize the weights
var weightSum = ImagesWeights.Last().Sum();
for (int i = 0; i < images.Length; i++)
{
ImagesWeights.Last()[i] /= weightSum;
}
// Train each weak classifier in parallel
weakClassifiers.AsParallel().ForAll(wc =>
{
wc.Train(ref integralImages, ref results, ImagesWeights.Last());
});
// Choosing weak classifier with lowest error
minFeature = weakClassifiers.OrderBy(wc => wc.Error).First();
if (double.IsNaN(minFeature.Error))
{
throw new Exception("Error is NaN");
}
// Saving weak classifier
WeakClassifiers.Add(minFeature);
weakClassifiers.Remove(minFeature);
// Calc Betas and Weights
Betas.Add(minFeature.Error / (1 - minFeature.Error));
Weights.Add(Math.Log(1.0 / Betas.Last()));
// Weights update
ImagesWeights.Add(new double[integralImages.Length]);
for (int i = 0; i < integralImages.Length; i++)
{
var featureResult = minFeature.Detect(integralImages[i]);
if (featureResult == results[i])
{
ImagesWeights.Last()[i] = ImagesWeights[ImagesWeights.Count - 2][i] * Betas[ImagesWeights.Count - 2];
}
else
{
ImagesWeights.Last()[i] = ImagesWeights[ImagesWeights.Count - 2][i];
}
}
var scores = CalcScores(integralImages, results);
DetermineThreshold(ref scores);
var detectionRate = 0.0;
var falsePositiveRate = 0.0;
Evaluate(integralImages, results, out falsePositiveRate, out detectionRate);
Console.WriteLine("Weak Classifier #{0} was added. Error: {1} Detection Rate: {2} False Positive Rate: {3}",
WeakClassifiers.Count, Math.Round(minFeature.Error, 5).ToString().PadRight(10),
Math.Round(detectionRate, 4).ToString().PadRight(10),
Math.Round(falsePositiveRate, 4).ToString().PadRight(10));
}