using System;
|
using System.Collections.Generic;
|
using System.Linq;
|
using System.Text;
|
using System.Threading.Tasks;
|
using Emgu.CV;
|
using Emgu.CV.CvEnum;
|
using Emgu.CV.ML;
|
using Emgu.CV.Structure;
|
|
namespace RichCreator.Editor.Utils
|
{
|
public class HOGTrain
|
{
|
|
public HOGTrain(System.Drawing.Size winSize,
|
System.Drawing.Size blockSize,
|
System.Drawing.Size strideSize,
|
System.Drawing.Size cellSize,
|
Int32 channel)
|
{
|
this.WinSize = winSize;
|
this.BlockSize = blockSize;
|
this.StrideSize = strideSize;
|
this.CellSize = cellSize;
|
this.Channel = channel;
|
|
Initizlize();
|
}
|
|
|
public System.Drawing.Size WinSize { get; set; }
|
|
public System.Drawing.Size BlockSize { get; set; }
|
|
public System.Drawing.Size StrideSize { get; set; }
|
|
public System.Drawing.Size CellSize { get; set; }
|
|
public Int32 Channel { get; set; }
|
|
/// <summary>
|
/// 训练数据
|
/// </summary>
|
private List<HogFeatureDataContainer> FeatureDatas = new List<HogFeatureDataContainer>();
|
|
/// <summary>
|
/// 维数
|
/// </summary>
|
public Int32 Dimensional { get; set; }
|
|
private HOGDescriptor hog;
|
|
private void Initizlize()
|
{
|
//计算Dimensional
|
/// a = (winsize宽-block宽) / stride宽 +1 23
|
/// b = (winsize高-block高) / stride高 +1 14
|
/// c = block宽 / cell宽 2
|
/// d = block高 / block高 2
|
/// 纬度 = a* b*c* d*channel数，本示例中:23*14*2*2*9=11592
|
Int32 a = (this.WinSize.Width - this.BlockSize.Width) / this.StrideSize.Width + 1;
|
Int32 b = (this.WinSize.Height - this.BlockSize.Height) / this.StrideSize.Height + 1;
|
Int32 c = this.BlockSize.Width / this.CellSize.Width;
|
Int32 d = this.BlockSize.Height / this.CellSize.Height;
|
this.Dimensional = a * b * c * d * Channel;
|
|
//第一个参数：winsize：窗口大小，我取样本大小；
|
//第二个参数：block，我的是12 * 12；当然，也可以设置为18 * 18或者6 * 6，切记：BLOCK大小不能是技术值，至于为什么，看一下cell和block的关系即可。同时：设计的过程中，BLOCK要可被winsize整除。网上的资料都是16 * 16，也没有说清楚为什么，其实这个看懂了HOG后可以自行设置
|
//第三个参数：stride，我的设置是6 * 6，当然也可以设置为12 * 12或者4 * 4，不同的设置，将得到不同的纬度数量；
|
//第四个参数：cell，4个cell构成1个block，这也就是block一定要是偶数的原因。
|
//第五个参数：需要得到几个channel的值，默认是9个，具体原因请参见HOG原论文
|
hog = new HOGDescriptor(this.WinSize, this.BlockSize, this.StrideSize, this.CellSize, this.Channel);
|
}
|
|
|
/// <summary>
|
/// 训练
|
/// </summary>
|
/// <param name="files">样本图片列表</param>
|
/// <param name="featureClasses">样本类型,1正向样本，-1负向样本</param>
|
public void Train(string[] files, Int32 featureClasses)
|
{
|
Mat tempMat = null;
|
Image<Gray, byte> im = null;
|
for (int f = 0; f < files.Length; f++)
|
{
|
tempMat = CvInvoke.Imread(files[f]);
|
im = tempMat.ToImage<Gray, byte>();
|
|
float[] fArr = hog.Compute(im);
|
this.FeatureDatas.Add(new HogFeatureDataContainer(fArr, featureClasses));
|
}
|
}
|
|
/// <summary>
|
///
|
/// </summary>
|
/// <param name="saveToXmlPath"></param>
|
public void Save(string saveToXmlPath)
|
{
|
Matrix<float> hogFeatureData = new Matrix<float>(this.FeatureDatas.Count, this.Dimensional);//样本数，HOG纬度
|
Matrix<int> featureClasses = new Matrix<int>(this.FeatureDatas.Count, 1);//样本数
|
|
HogFeatureDataContainer container;
|
for (int i = 0; i < this.FeatureDatas.Count; i++)
|
{
|
container = this.FeatureDatas[i];
|
for (int t = 0; t < container.FeatureDatas.Length; t++)
|
{
|
hogFeatureData[i, t] = container.FeatureDatas[t];
|
}
|
featureClasses.Data[i, 0] = container.FeatureClass;
|
}
|
|
//训练并保存训练结果
|
SVM svm = new SVM();
|
svm.Type = SVM.SvmType.CSvc;
|
svm.SetKernel(SVM.SvmKernelType.Linear);//线性
|
svm.C = 1;
|
svm.TermCriteria = new MCvTermCriteria(1000, 0.001);//1000次或者收敛达到0.001就跳出
|
svm.Train(hogFeatureData, Emgu.CV.ML.MlEnum.DataLayoutType.RowSample, featureClasses);
|
svm.Save(saveToXmlPath);
|
}
|
|
|
/// <summary>
|
/// 训练数据包装器
|
/// </summary>
|
private struct HogFeatureDataContainer
|
{
|
public HogFeatureDataContainer(float[] featureDatas, Int32 featureClass)
|
{
|
this.FeatureDatas = featureDatas;
|
this.FeatureClass = featureClass;
|
}
|
|
public float[] FeatureDatas;
|
|
public Int32 FeatureClass;
|
}
|
}
|
}
|
