private SampleCollection CreateNextSamples(SampleCollection allNegSamples, double falsePositiveRate, int maxSampleNum)
{
int selectSampleCount = (int)(maxSampleNum / falsePositiveRate);
if (selectSampleCount <= 0)
{
selectSampleCount = allNegSamples.Count;
}
SampleCollection negSamples = allNegSamples.GetNegSamples(selectSampleCount);
int oldNum = negSamples.Count;
negSamples = this.GetPositivePredictedSamples(negSamples);
if (DebugMsg.Debug)
{
string msg = string.Format("选取了{0}个负样本，通过了{1}个，误检率为{2:P5}\r\n",
oldNum, negSamples.Count, (MyFloat)negSamples.Count / oldNum);
DebugMsg.AddMessage(msg, 0);
}
if (negSamples.Count > maxSampleNum)
{
negSamples.TrimExcess(maxSampleNum);
}
GC.Collect();
return(negSamples);
}
void worker_RunWorkerCompleted(object sender, RunWorkerCompletedEventArgs e)
{
try
{
File.WriteAllText(@"D:\ccc.txt", textBox_debug.Text, Encoding.Default);
}
catch (System.Exception ex)
{
DebugMsg.AddMessage(ex.ToString(), 0);
}
}
private void button_train_Click(object sender, EventArgs e)
{
textBox_debug.BringToFront();
if (_cascadeClassifier != null)
{
if (MessageBox.Show("已存在分类器，在现有分类器上继续训练吗？", "", MessageBoxButtons.YesNo) == DialogResult.No)
{
_cascadeClassifier = new CascadeClassifier();
}
}
else
{
_cascadeClassifier = new CascadeClassifier();
}
if (!inited)
{
button_loadSamples_Click(sender, e);
// WeakClassifierManager.Instance.CreateHaarFeatures(_size.Width, _size.Height , _colorType);
WeakClassifierManager.Instance.CreateHaarFeatures(_size.Width, _size.Height / 2, _colorType);
WeakClassifierManager.Instance.AddSymmetricHaarFeatures(_size.Width, _size.Height, _colorType);
}
//StageClassifier.ViewId = int.Parse(textBox1.Text);
GC.Collect();
GC.WaitForFullGCComplete();
BackgroundWorker worker = new BackgroundWorker();
worker.DoWork += new DoWorkEventHandler(
(a, b) =>
{
//cascadeClassifier.Train(posSamples, negSamples, (MyFloat)0.001, (MyFloat)0.5, (MyFloat)0.99);
_cascadeClassifier.Train(_posSamples, _negSamples, _validateSamples, _size, _targetFalsePositiveRate, _maxFalsePositiveRate, _minHitRate);
});
worker.ProgressChanged += new ProgressChangedEventHandler(worker_ProgressChanged);
worker.RunWorkerCompleted += new RunWorkerCompletedEventHandler(worker_RunWorkerCompleted);
worker.WorkerReportsProgress = true;
DebugMsg.Init(worker);
worker.RunWorkerAsync();
inited = true;
}
public void Train(SampleCollection posSamples,
SampleCollection allNegSamples,
SampleCollection validateSamples,
Size size,
double targetFalsePositiveRate, double maxFalsePositiveRate, double minHitRate)
{
const int startCapacity = 20;
_size = size;
List <StageClassifier> stageClassifier = new List <StageClassifier>(startCapacity);
double hitRate = 1;
double falsePositiveRate = 1;
int stageCount = 0;
int maxSampleNum;
SampleCollection negSamples;
{ //计算使用的负样本数量
double MaxMemoryUse = MemoryInfo.GetFreePhysicalMemory() - 400e6;
if (MaxMemoryUse < 0)
{
MaxMemoryUse = 1.8e9;
}
//MaxMemoryUse = 2.48e9;
//按照内存大小来选取负样本，在充分利用内存的同时防止使用虚拟内存
int weakNum = WeakClassifierManager.Instance.WeakClassifiers.Length;
maxSampleNum = (int)((
MaxMemoryUse
- 1.5e6f //预留150M给程序其它部分
)
/ weakNum / sizeof(MyFloat)) - posSamples.Count;
if (DebugMsg.Debug)
{
string msg = string.Format("\r\n单次使用负样本：{0}，特征数量：{1}，预计消耗内存：{2}M\r\n",
maxSampleNum, weakNum, MaxMemoryUse / 1024 / 1024);
DebugMsg.AddMessage(msg, 0);
}
}
//如果已有分类器，计算当前的性能
if (_classifiers != null && _classifiers.Length > 0)
{
PredictResult result = EvaluateErrorRate(validateSamples);
hitRate = result.HitRate;
falsePositiveRate = result.FalsePositiveRate;
stageCount = _classifiers.Length;
stageClassifier.AddRange(_classifiers);
negSamples = CreateNextSamples(allNegSamples, falsePositiveRate, maxSampleNum);
if (DebugMsg.Debug)
{
string msg = string.Format("载入分类器完成，层数：{0}，当前性能为：\r\n检测率：\t{1:P5}，\t误检率：\t{2:P5}\r\n------\r\n",
stageCount, hitRate, falsePositiveRate);
DebugMsg.AddMessage(msg, 0);
}
}
else
{
negSamples = allNegSamples.GetNegSamples(maxSampleNum);
}
while (falsePositiveRate > targetFalsePositiveRate && negSamples.Count != 0 && posSamples.Count != 0)
{
stageCount++;
if (DebugMsg.Debug)
{
string msg = string.Format("--------------------\r\n开始训练第{0}级分类器，使用的数量为：\r\n正样本：\t{1}\t负样本：\t{2}\r\n目标检测率：\t{3:P5}\t目标误检率：\t{4:P5}\r\n",
stageCount, posSamples.Count, negSamples.Count, minHitRate, maxFalsePositiveRate);
DebugMsg.AddMessage(msg, 0);
}
StageClassifier currentStage = new StageClassifier(stageCount);
PredictResult result = currentStage.Train(posSamples, negSamples, validateSamples,
maxFalsePositiveRate, minHitRate);
falsePositiveRate *= result.FalsePositiveRate;
hitRate *= result.HitRate;
stageClassifier.Add(currentStage);
_classifiers = stageClassifier.ToArray();
if (DebugMsg.Debug)
{
string msg = string.Format("------\r\n第{0}级分类器训练结束，当前性能为：\r\n检测率：\t{1:P5}，\t误检率：\t{2:P5}\r\n弱分类器数量：{3}\r\n目前训练时间总计：{4}\r\n",
stageCount, hitRate, falsePositiveRate, currentStage.WeakClassifierCount, DebugMsg.stopwatch.Elapsed.ToString());
DebugMsg.AddMessage(msg, 0);
}
if (falsePositiveRate > targetFalsePositiveRate)
{
//posSamples = this.CreateNextSamples(posSamples);
validateSamples = this.GetPositivePredictedSamples(validateSamples);
negSamples = CreateNextSamples(allNegSamples, falsePositiveRate, maxSampleNum);
}
this.Save(string.Format(@"D:\ccc{0}.xml", stageCount.ToString()));
}
if (DebugMsg.Debug)
{
DebugMsg.stopwatch.Stop();
string msg = string.Format("\r\n--------------------\r\n全部训练完成，训练时间总计：{0}\r\n",
DebugMsg.stopwatch.Elapsed.ToString());
DebugMsg.AddMessage(msg, 0);
}
this.Save(@"D:\ccc.xml");
}
/// <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);
}
