public object Clone()
{
WeakClassifier newWeak = (WeakClassifier)this.MemberwiseClone();
newWeak._featureValues = (MyFloat[])this._featureValues.Clone();
return(newWeak);
}
internal static WeakClassifier Load(float threshold, bool posLargeThanThreshold, float weight, System.Drawing.Rectangle[] rects, int[] weights, ColorType colorType)
{
HaarFeature feature = new HaarFeature(0, rects, weights, colorType);
WeakClassifier weak = new WeakClassifier(feature);
weak._threshold = threshold;
weak._posLargeThanThreshold = posLargeThanThreshold;
weak._weight = weight;
return(weak);
}
private void UpdateWeights(WeakClassifier bestClassifier, int numPos, int numNeg, float[] sampleWeight)
{
int count = numNeg + numPos;
double err = 0;
int i;
bool[] classifyResult = new bool[count];
bool result;
for (i = 0; i < numPos; i++)
{
result = bestClassifier.Predict(i);
if (result == false)
{
err += sampleWeight[i]; //累加分类错误的样本权值
}
classifyResult[i] = result;
}
for (i = numPos; i < count; i++)
{
result = bestClassifier.Predict(i);
if (result == true)
{
err += sampleWeight[i]; //累加分类错误的样本权值
}
classifyResult[i] = !result;
}
//调整样本权值
// logRatio( double val )
// {
// const double eps = 1e-5;
//
// val = max( val, eps );
// val = min( val, 1. - eps );
// return log( val/(1. - val) );
// }
const double eps = (double)1e-5;
double val = err;
val = Math.Max(val, eps);
val = Math.Min(val, 1 - eps);
err = val;
MyFloat factor = (MyFloat)(err / (1 - err));
for (int indexSample = 0; indexSample < count; indexSample++)
{
if (classifyResult[indexSample])
{
sampleWeight[indexSample] *= factor;
}
}
bestClassifier.Weight = -(MyFloat)Math.Log10(factor); //设定弱分类器权值α
}
private static StageClassifier LoadStageClassifier(int id, XmlNode stageNode)
{
MyFloat threshold = MyFloat.Parse(stageNode.Attributes["Threshold"].Value);
int weakNum = int.Parse(stageNode.Attributes["ClassifiersNum"].Value);
WeakClassifier[] weaks = new WeakClassifier[weakNum];
for (int i = 0; i < weakNum; i++)
{
XmlNode weakNode = stageNode.ChildNodes[i];
weaks[i] = LoadWeakClassifier(weakNode);
}
StageClassifier stage = StageClassifier.Load(id, threshold, weaks);
return(stage);
}
private static WeakClassifier LoadWeakClassifier(XmlNode weakNode)
{
MyFloat threshold = MyFloat.Parse(weakNode.Attributes["Threshold"].Value);
MyFloat weight = MyFloat.Parse(weakNode.Attributes["Weight"].Value);
bool posLargeThanThreshold = bool.Parse(weakNode.Attributes["PosLargeThanThreshold"].Value);
ColorType colorType = ColorType.Null;
if (weakNode.Attributes["ColorType"] != null)
{
string colorValue = weakNode.Attributes["ColorType"].Value;
if (colorValue.Contains(ColorType.Gray.ToString()))
{
colorType |= ColorType.Gray;
}
if (colorValue.Contains(ColorType.Saturation.ToString()))
{
colorType |= ColorType.Saturation;
}
if (colorType == ColorType.Null)
{
colorType = ColorType.Gray;
}
}
else
{
colorType = ColorType.Gray;
}
List <Rectangle> rects = new List <Rectangle>(4);
List <int> weights = new List <int>(4);
foreach (XmlNode node in weakNode.ChildNodes)
{
string[] nums = node.InnerText.Split(' ');
int x = int.Parse(nums[0]);
int y = int.Parse(nums[1]);
int width = int.Parse(nums[2]);
int height = int.Parse(nums[3]);
int ww = int.Parse(nums[4]);
rects.Add(new Rectangle(x, y, width, height));
weights.Add(ww);
}
return(WeakClassifier.Load(threshold, posLargeThanThreshold, weight, rects.ToArray(), weights.ToArray(), colorType));
}
private void AddRect(int x1, int y1, int w1, int h1, int ww1, int x2, int y2, int w2, int h2, int ww2)
{
Rectangle[] rects = new Rectangle[] {
new Rectangle(x1, y1, w1, h1),
new Rectangle(x2, y2, w2, h2),
};
int[] weights = new int[] { ww1, ww2 };
if (_gray)
{
HaarFeature gray = new HaarFeature(_weakId, rects, weights, ColorType.Gray);
_weakClassifiers[_weakId] = new WeakClassifier(gray);
_weakId++;
}
if (_saturation)
{
HaarFeature saturation = new HaarFeature(_weakId, rects, weights, ColorType.Saturation);
_weakClassifiers[_weakId] = new WeakClassifier(saturation);
_weakId++;
}
}
private void treeView1_AfterSelect(object sender, TreeViewEventArgs e)
{
if ((e.Node.Tag is WeakClassifier) == false)
{
return;
}
WeakClassifier weak = e.Node.Tag as WeakClassifier;
textBox1.Text = weak.GetNumString();
Image <Gray, Byte> imgBlank = _imgBlank.Copy();
Image <Bgr, Byte> imgPos = _imgPos.Copy();
HaarFeature feature = (HaarFeature)weak.Feature;
for (int i = 0; i < feature._rects.Length; i++)
{
Image <Gray, Byte> mask = new Image <Gray, Byte>(imgBlank.Size);
int weight = feature._weights[i];
Rectangle rect = feature._rects[i];
rect.Height--;
rect.Width--;
mask.Draw(rect, new Gray(Math.Abs(weight)), 0);
if (weight > 0)
{
imgBlank = imgBlank.Add(mask);
}
else
{
imgBlank = imgBlank.Sub(mask);
}
imgPos.Draw(rect, new Bgr(), 1);
}
Image <Gray, Byte> light = imgBlank.ThresholdBinary(new Gray(128), new Gray(255));
Image <Gray, Byte> dark = imgBlank.ThresholdBinaryInv(new Gray(127), new Gray(255));
imgBlank.SetValue(new Gray(255), light);
imgBlank.SetValue(new Gray(0), dark);
imageBox1.Image = imgBlank;
imageBox2.Image = imgPos;
}
/// <summary>
/// 使用AdaBoost算法训练一次，得到一个最优的弱分类器
/// </summary>
/// <param name="posSamples">正样本集</param>
/// <param name="negSamples">负样本集</param>
/// <param name="sampleWeight">样本权值，依次包含了正样本和负样本的权值</param>
/// <returns>分类误差最优的弱分类器</returns>
private WeakClassifier AdaBoost(int numPos, int numNeg, MyFloat[] sampleWeight)
{
int count = numNeg + numPos;
double minErr = double.MaxValue;
int minErrIndex = 0, i;
WeakClassifier[] weakClassifiers = WeakClassifierManager.Instance.WeakClassifiers;
int classifierCount = weakClassifiers.Length;
for (int indexClassifier = 0; indexClassifier < classifierCount; indexClassifier++)
{
WeakClassifier classifier = weakClassifiers[indexClassifier];
double errCount = 0;
bool result;
for (i = 0; i < numPos; i++)
{
//result=classifier.Predict(posSamples[i]);
result = classifier.Predict(i);
if (result == false)
{
errCount += sampleWeight[i]; //累加分类错误的样本权值
}
}
for (; i < count; i++)
{
//result = classifier.Predict(negSamples[i]);
result = classifier.Predict(i);
if (result == true)
{
errCount += sampleWeight[i]; //累加分类错误的样本权值
}
}
if (errCount < minErr) //记录最佳弱分类器
{
minErr = errCount;
minErrIndex = indexClassifier;
}
}
//DebugMsg.AddMessage(minErr.ToString(), 0);
bool[] classifyResult = new bool[count];
WeakClassifier bestClassifier = (WeakClassifier)weakClassifiers[minErrIndex].Clone();
{
bool result;
for (i = 0; i < numPos; i++)
{
result = bestClassifier.Predict(i);
classifyResult[i] = result;
}
for (; i < count; i++)
{
result = bestClassifier.Predict(i);
classifyResult[i] = !result;
}
}
//调整样本权值
// logRatio( double val )
// {
// const double eps = 1e-5;
//
// val = max( val, eps );
// val = min( val, 1. - eps );
// return log( val/(1. - val) );
// }
const double eps = (double)1e-5;
double val = minErr;
val = Math.Max(val, eps);
val = Math.Min(val, 1 - eps);
minErr = val;
MyFloat factor = (MyFloat)(minErr / (1 - minErr));
for (int indexSample = 0; indexSample < count; indexSample++)
{
if (classifyResult[indexSample])
{
sampleWeight[indexSample] *= factor;
}
}
//WeakClassifier bestClassifier= (WeakClassifier)weakClassifiers[minErrIndex].Clone();
//bestClassifier.Weight =(MyFloat) Math.Log10(1 / factor); //设定弱分类器权值α
bestClassifier.Weight = -(MyFloat)Math.Log10(factor); //设定弱分类器权值α
return(bestClassifier);
}
/// <summary>
/// 训练级联分类器中一层的强分类器
/// </summary>
/// <param name="posSamples"></param>
/// <param name="negSamples"></param>
/// <param name="maxFalsePositiveRate"></param>
/// <param name="minHitRate"></param>
/// <returns>训练的统计结果</returns>
public PredictResult Train(SampleCollection posSamples,
SampleCollection negSamples,
SampleCollection validateSamples,
double maxFalsePositiveRate,
double minHitRate)
{
List <WeakClassifier> weakClassifiers = new List <WeakClassifier>(10);
PredictResult result = new PredictResult();
MyFloat[] sampleWeight = InitWeight(posSamples.Count, negSamples.Count);
WeakClassifierManager allWeak = WeakClassifierManager.Instance;
Stopwatch watch = new Stopwatch();
watch.Start();
allWeak.PreCalcFeatureValue(posSamples, negSamples);
watch.Stop();
if (DebugMsg.Debug)
{
string msg = string.Format("所有弱分类器特征值预计算完成，用时：{0}\r\n",
watch.Elapsed.ToString());
DebugMsg.AddMessage(msg, 0);
}
int trainTime = 0;
do
{
if (++trainTime != 1)
{
NormalizeWeight(sampleWeight);
}
if (DebugMsg.Debug)
{
string msg = string.Format("开始训练第{0}个弱分类器\r\n",
trainTime);
DebugMsg.AddMessage(msg, 0);
}
watch.Reset();
watch.Start();
allWeak.Train(posSamples, negSamples, sampleWeight);
WeakClassifier newBestClassifier = AdaBoost(posSamples.Count, negSamples.Count, sampleWeight);
//UpdateWeights(newBestClassifier, posSamples.Count, negSamples.Count, sampleWeight);
weakClassifiers.Add(newBestClassifier);
_classifiers = weakClassifiers.ToArray();
result = EvaluateErrorRate(validateSamples, minHitRate, maxFalsePositiveRate);
watch.Stop();
if (DebugMsg.Debug)
{
string msg = string.Format("训练完成，花费时间{0}\r\n检测率：\t{1:P5}\t误检率：\t{2:P5}\r\n",
watch.Elapsed.ToString(),
result.HitRate,
result.FalsePositiveRate);
DebugMsg.AddMessage(msg, 1);
}
} while (result.FalsePositiveRate > maxFalsePositiveRate);
allWeak.ReleaseTrainData();
foreach (WeakClassifier weak in _classifiers)
{
weak.ReleaseTrainData();
}
return(result);
}