/// <summary>
/// 构造函数，由Hog图生成块内归一化后的向量图
/// </summary>
/// <param name="hogGram">Hog图</param>
/// <param name="blockWidth">块宽</param>
/// <param name="blockHeight">块高</param>
public NormBlockVectorGram(HogGram hogGram, int blockWidth, int blockHeight)
{
BlockSize.Width = blockWidth;
BlockSize.Height = blockHeight;
int hogWidth = hogGram.HogSize.Width;
int hogHeight = hogGram.HogSize.Height;
int hogPartNumber = hogGram.PartNumber;
BlockGramWidth = hogWidth - blockWidth + 1;
BlockGramHeight = hogHeight - blockHeight + 1;
NormBlockVectors = new ArrayList();
if (hogGram != null && hogWidth >= blockWidth && hogHeight >= blockWidth)
{
HGram = hogGram;
for (int row = 0; row < BlockGramHeight; row++)
{
for (int col = 0; col < BlockGramWidth; col++)
{
double[] vec = new double[blockWidth * blockHeight * hogPartNumber];
double vecsum = 0;
for (int i = 0; i < blockHeight; i++)
{
for (int j = 0; j < blockWidth; j++)
{
for (int p = 0; p < hogPartNumber; p++)
{
double r = hogGram.HogCells[(col + j) + (row + i) * hogWidth].HistElements[p].rho;
// hogGram.HogCells的索引是从小到大。
vec[p + j * hogPartNumber + i * hogPartNumber * blockWidth] = r;
//vecsum += r * r; // 这种归一化是将向量约束在“半径”为1的“球面”上，因为r>=0，实际上是分布在1/4个球面
vecsum += r; // 这种归一化是将向量约束在“边长为1的”立方体“内
}
}
}
// 归一化向量（非常重要）
for (int i = 0; i < vec.Length; i++)
{
if (vecsum != 0)
{
//vec[i] = vec[i] / Math.Sqrt(vecsum); // 对应上述第一种归一化法
vec[i] = vec[i] / vecsum;
}
else
{
vec[i] = 0;
}
}
NormBlockVectors.Add(vec);
}
}
}
}
/// <summary>
/// 绘制Hog示意图
/// </summary>
/// <param name="hogGram">Hog图数据</param>
/// <param name="width">示意图宽</param>
/// <param name="height">示意图高</param>
public static Bitmap DrawHogGram(HogGram hogGram, int width, int height)
{
Bitmap bmp = new Bitmap(width, height, PixelFormat.Format24bppRgb);
Graphics g = Graphics.FromImage(bmp);
int grid_width = width / hogGram.HogSize.Width; // 格子宽
int grid_height = height / hogGram.HogSize.Height; // 格子高
// 格子中心相对坐标
PointF rel_center = new PointF((float)(grid_width / 2), (float)(grid_height / 2));
float radius = Math.Min((float)(grid_width / 2), (float)(grid_height / 2));
double maxval = 0;
double[] transval = new double[hogGram.HogCells.Length * hogGram.PartNumber];
for (int i = 0; i < hogGram.HogCells.Length; i++)
{
for (int j = 0; j < hogGram.PartNumber; j++)
{
transval[j + i * hogGram.PartNumber] = hogGram.HogCells[i].HistElements[j].rho;
if (transval[j + i * hogGram.PartNumber] > maxval)
{
maxval = transval[j + i * hogGram.PartNumber];
}
}
}
for (int i = 0; i < transval.Length; i++)
{
if (maxval != 0)
{
transval[i] = 255 * Math.Sqrt(transval[i]) / Math.Sqrt(maxval);
}
}
HogHistElement he;
PointF center;
PointF right;
PointF left;
for (int row = 0; row < hogGram.HogSize.Height; row++)
{
for (int col = 0; col < hogGram.HogSize.Width; col++)
{
for (int i = 0; i < hogGram.PartNumber; i++)
{
he = hogGram.HogCells[col + row * hogGram.HogSize.Width].HistElements[i];
center = new PointF(col * grid_width + grid_width / 2, row * grid_height + grid_height / 2);
right = new PointF((float)(center.X + radius * Math.Cos(he.theta)), (float)(center.Y + radius * Math.Sin(he.theta)));
left = new PointF((float)(center.X - radius * Math.Cos(he.theta)), (float)(center.Y - radius * Math.Sin(he.theta)));
int idx = i + (col + row * hogGram.HogSize.Width) * hogGram.PartNumber;
// hog示意图体现在某个方向灰度的亮度上，而不是在该方向上线段的长短上
Pen p = new Pen(Color.FromArgb((int)transval[idx], (int)transval[idx], (int)transval[idx]));
g.DrawLine(p, right, left);
}
}
}
//Pen p = new Pen(Color.Red);
//g.DrawLine(p, new PointF(20.5f, 20.5f), new PointF(100.4f, 100.4f));
return(bmp);
}
/// <summary>
/// 从位图（灰度图）中获取Hog图，位图像素是以从左到右、从下到上的顺序在内存中从低地址排列到高地址
/// 但是C#的Bitmap对象好像把这顺序颠倒过来了，根据实际经验发现，在BitmapData中，位图数据以从左到
/// 右、从上到下的顺序排列，和图像坐标的顺序一致
/// </summary>
/// <param name="bitmapSource">源位图</param>
/// <param name="cellHeight">单元的高度</param>
/// <param name="cellWidth">单元的宽度</param>
/// <param name="partNumber">PI弧度等分数</param>
/// <returns>Hog图</returns>
public static HogGram GetHogFromBitmap(Bitmap bitmapSource, int cellWidth, int cellHeight, int partNumber)
{
HogGram hogGram = null;
if (bitmapSource != null && bitmapSource.PixelFormat == PixelFormat.Format8bppIndexed)
{
int width = bitmapSource.Width;
int height = bitmapSource.Height;
hogGram = new HogGram(width, height, cellWidth, cellHeight, partNumber);
if (hogGram == null)
{
return(hogGram);
}
Rectangle rect = new Rectangle(0, 0, width, height);
// 获得位图内容数据
BitmapData dataSource = bitmapSource.LockBits(rect, ImageLockMode.ReadOnly, bitmapSource.PixelFormat);
// Stride为位图中每一行以4字节对齐的行宽
int strideSource = dataSource.Stride;
unsafe
{
byte *ptrSource = (byte *)dataSource.Scan0.ToPointer();
byte *ptr1 = null;
for (int row = 0; row < height; row++)
{
ptr1 = ptrSource + strideSource * row;
for (int col = 0; col < width; col++)
{
GradientElement ge = new GradientElement();
if (row == 0 || row == height - 1 || col == 0 || col == width - 1)
{
ge.rho = 0;
ge.theta = 0;
}
else
{
// 位图的在内存中的排列是从左到右，从下到上的，但BitmapData貌似做了优化，把位图数据在内存中的排列
// 顺序改为与坐标系一致，即从左到右、从上到下
double gradX = *(ptr1 + 1) - *(ptr1 - 1);
double gradY = *(ptr1 + strideSource) - *(ptr1 - strideSource);
ge.rho = Math.Sqrt(gradX * gradX + gradY * gradY);
ge.theta = Math.Atan2(gradY, gradX); // 注意坐标系是垂直翻转的
}
int bmpX = col;
int bmpY = row;
//int bmpY = height - 1 - row; // 由于BitmapData自动颠倒了位图存储序，所以这个写法是错的。
hogGram.VoteHog(bmpX, bmpY, ge);
ptr1++;
}
}
}
bitmapSource.UnlockBits(dataSource);
}
return(hogGram);
}
/// <summary>
/// 训练正样本
/// </summary>
/// <param name="bmp">正样本位图</param>
public void TrainPositive(Bitmap bmp)
{
Bitmap samplebmp = null;
double neg_distance = 0;
double pos_distance = 0;
bool hasinserted = false; // 指明样本是否已插入队列
samplebmp = ImgOper.ResizeImage(bmp, Parameter.DETECT_WINDOW_SIZE.Width, Parameter.DETECT_WINDOW_SIZE.Height);
samplebmp = ImgOper.Grayscale(samplebmp);
for (double angle = (-1) * Parameter.ANGLE_BORDER; angle < Parameter.ANGLE_BORDER; angle += Parameter.ANGLE_INTERVAL)
{
Bitmap bmpclone = ImgOper.RotateImage(samplebmp, angle);
bmpclone = ImgOper.ResizeImage(bmpclone, Parameter.DETECT_WINDOW_SIZE.Width, Parameter.DETECT_WINDOW_SIZE.Height);
for (double scale = (-1) * Parameter.SCALE_BORDER; scale < Parameter.SCALE_BORDER; scale += Parameter.SCALE_INTERVAL)
{
// 往两个方向去，所以是减号
IntPoint lt = new IntPoint((int)(bmpclone.Width * scale / 2), (int)(bmpclone.Height * scale / 2));
IntPoint rt = new IntPoint(bmpclone.Width - 1 - (int)(bmpclone.Width * scale / 2), (int)(bmpclone.Height * scale / 2));
IntPoint rb = new IntPoint(bmpclone.Width - 1 - (int)(bmpclone.Width * scale / 2), bmpclone.Height - 1 - (int)(bmpclone.Height * scale / 2));
IntPoint lb = new IntPoint((int)(bmpclone.Width * scale / 2), bmpclone.Height - 1 - (int)(bmpclone.Height * scale / 2));
Bitmap scalebmp = ImgOper.QuadrilateralTransform(bmpclone, lt, rt, rb, lb);
HogGram hogGram = HogGram.GetHogFromBitmap(scalebmp, Parameter.CELL_SIZE.Width, Parameter.CELL_SIZE.Height, Parameter.PART_NUMBER);
NormBlockVectorGram blockGram = new NormBlockVectorGram(hogGram, Parameter.BLOCK_SIZE.Width, Parameter.BLOCK_SIZE.Height);
Rectangle rect = new Rectangle(0, 0, hogGram.HogSize.Width, hogGram.HogSize.Height);
double[] vect = blockGram.GetHogWindowVec(rect);
if (Dimension != 0 && vect.Length != Dimension)
{
throw new Exception("输入正样本的尺寸与其他样本尺寸不一致！");
}
ValuedBitmap vbmp = null;
if (NegCenter != null && PosCenter != null)
{
// 计算离正负中心的距离
for (int i = 0; i < vect.Length; i++)
{
neg_distance += Math.Abs(vect[i] - NegCenter[i]);
pos_distance += Math.Abs(vect[i] - PosCenter[i]);
}
// 与负样本中心重合时，说明是负样本，不能插入正样本队列
if (neg_distance == 0)
{
return;
}
// 检测到的正样本加入样本队列的第二道关，如果不够接近正样本中心，就无法加入队列
// 按照Hog检测的判定条件，正距离乘以Parameter.POS_DIST_COEF，使其避开边界
if (neg_distance < pos_distance * Parameter.POS_DIST_COEF)
{
return;
}
// 带归一化的系数，如果用pos_distance/neg_distance，值可能会溢出;
// 将pos_distance / (pos_distance + neg_distance)作为正样本的评价系数，值越小越接近正样本
vbmp = new ValuedBitmap(scalebmp, pos_distance / (pos_distance + neg_distance));
}
else
{
// 如果正或负样本库还没建立起来，则Val暂时赋值为1
vbmp = new ValuedBitmap(scalebmp, 1);
}
// 检测到的正样本加入样本队列的第三道关，与正样本评价系数的有序队列比较后，决定是否加入样本队列
hasinserted = InsertValuedBitmap(ref PosMapCollection, vbmp, Parameter.POS_LIMITED_NUMBER);
PosLength = PosMapCollection.Count;
//// 人工观察正样本插入情况
//if (hasinserted && vbmp != null)
//{
// vbmp.VBitmap.Save("Image\\pos_save\\" + poscnt + "_" + vbmp.Val + ".jpg");
// poscnt++;
//}
// 如果样本已经插入队列，说明样本比较可信，重新计算样本中心
if (hasinserted)
{
if (PosCenter == null)
{
Dimension = vect.Length;
PosCenter = new double[Dimension];
}
for (int i = 0; i < Dimension; i++)
{
PosCenter[i] = (PosCenter[i] * PosLength + vect[i]) / (PosLength + 1);
}
}
}
}
}
/// <summary>
/// 计算最近正样本距离系数，按照距离而不是相关系数，这样效率高，系数越小，检测对象越接近目标
/// </summary>
/// <param name="rect">检测目标区域</param>
/// <param name="bmp">位图</param>
/// <returns>最近距离系数， double.MaxValue表示计算异常或没计算</returns>
private double NearestNeighbour(Rectangle rect, Bitmap bmp)
{
Bitmap sample = null;
Bitmap detect = null;
Rectangle gramRect = Rectangle.Empty;
HogGram sampleHGram = null;
HogGram detectHGram = null;
NormBlockVectorGram sampleBlock = null;
NormBlockVectorGram detectBlock = null;
double[] detectvect = null;
double[] samplevect = null;
ArrayList posvects = null;
ArrayList negvects = null;
double minposdist = double.MaxValue;
double minnegdist = double.MaxValue;
double dist = 0;
double nearestdist = double.MaxValue;
if (PosLength == 0 || NegLength == 0)
{
return(nearestdist);
}
// 正样本载入
posvects = new ArrayList();
for (int i = 0; i < PosLength; i++)
{
sample = PosMapCollection[i].VBitmap;
sample = ImgOper.ResizeImage(sample, Parameter.DETECT_WINDOW_SIZE.Width, Parameter.DETECT_WINDOW_SIZE.Height);
sample = ImgOper.Grayscale(sample);
sampleHGram = HogGram.GetHogFromBitmap(sample, Parameter.CELL_SIZE.Width, Parameter.CELL_SIZE.Height, Parameter.PART_NUMBER);
sampleBlock = new NormBlockVectorGram(sampleHGram, Parameter.BLOCK_SIZE.Width, Parameter.BLOCK_SIZE.Height);
gramRect = new Rectangle(0, 0, sampleHGram.HogSize.Width, sampleHGram.HogSize.Height);
samplevect = sampleBlock.GetHogWindowVec(gramRect);
posvects.Add(samplevect);
}
// 负样本载入
negvects = new ArrayList();
for (int i = 0; i < NegLength; i++)
{
sample = NegMapCollection[i].VBitmap;
sample = ImgOper.ResizeImage(sample, Parameter.DETECT_WINDOW_SIZE.Width, Parameter.DETECT_WINDOW_SIZE.Height);
sample = ImgOper.Grayscale(sample);
sampleHGram = HogGram.GetHogFromBitmap(sample, Parameter.CELL_SIZE.Width, Parameter.CELL_SIZE.Height, Parameter.PART_NUMBER);
sampleBlock = new NormBlockVectorGram(sampleHGram, Parameter.BLOCK_SIZE.Width, Parameter.BLOCK_SIZE.Height);
gramRect = new Rectangle(0, 0, sampleHGram.HogSize.Width, sampleHGram.HogSize.Height);
samplevect = sampleBlock.GetHogWindowVec(gramRect);
negvects.Add(samplevect);
}
detect = ImgOper.CutImage(bmp, rect.X, rect.Y, rect.Width, rect.Height);
detect = ImgOper.ResizeImage(detect, Parameter.DETECT_WINDOW_SIZE.Width, Parameter.DETECT_WINDOW_SIZE.Height);
detect = ImgOper.Grayscale(detect);
detectHGram = HogGram.GetHogFromBitmap(detect, Parameter.CELL_SIZE.Width, Parameter.CELL_SIZE.Height, Parameter.PART_NUMBER);
detectBlock = new NormBlockVectorGram(detectHGram, Parameter.BLOCK_SIZE.Width, Parameter.BLOCK_SIZE.Height);
gramRect = new Rectangle(0, 0, detectHGram.HogSize.Width, detectHGram.HogSize.Height);
detectvect = detectBlock.GetHogWindowVec(gramRect);
foreach (double[] svect in posvects)
{
dist = ImgStatCompute.ComputeDistance(detectvect, svect);
if (dist < minposdist)
{
minposdist = dist;
}
}
foreach (double[] svect in negvects)
{
dist = ImgStatCompute.ComputeDistance(detectvect, svect);
if (dist < minnegdist)
{
minnegdist = dist;
}
}
if (minnegdist != 0 || minposdist != 0)
{
nearestdist = minposdist / (minposdist + minnegdist);
}
return(nearestdist);
}
/// <summary>
/// Hog检测, 被检测到图像自动缩放到BMPLIMITSIZE容忍范围内，并在检测完后将检测框自动放大之前缩小的倍率
/// </summary>
/// <param name="bmp">位图</param>
public RectangleCollection HogDetect(Bitmap bmp)
{
RectangleCollection resultCollection = null;
if (bmp == null)
{
return(null);
}
if (NegCenter == null && PosCenter == null)
{
return(null);
}
DateTime dt = DateTime.Now;
double elapse = 0;
// 针对原图的缩放倍率
double se = 1;
if (bmp.Width > Parameter.BMPLIMITSIZE.Width || bmp.Height > Parameter.BMPLIMITSIZE.Height)
{
se = bmp.Width / (double)Parameter.BMPLIMITSIZE.Width > bmp.Height / (double)Parameter.BMPLIMITSIZE.Height ?
bmp.Width / (double)Parameter.BMPLIMITSIZE.Width : bmp.Height / (double)Parameter.BMPLIMITSIZE.Height;
bmp = ImgOper.ResizeImage(bmp, (int)(bmp.Width / se), (int)(bmp.Height / se));
}
bmp = ImgOper.Grayscale(bmp);
//bmp = ImgOper.GaussianConvolution(bmp, GAUSSIAN_SIGMA, GAUSSIAN_SIZE); // 高斯卷积，使得图像平滑
// 所有层的检测结果
ArrayList resultlayers = new ArrayList();
// 初始缩放因子
double scalecoef = 1.0;
Bitmap scalebmp = null;
int newwidth = (int)(bmp.Width / scalecoef);
int newheight = (int)(bmp.Height / scalecoef);
// 每层最小距离点的集合
ArrayList idx_layermindistance = new ArrayList();
int cnt = 0;
do
{
scalebmp = ImgOper.ResizeImage(bmp, newwidth, newheight);
HogGram hogGram = HogGram.GetHogFromBitmap(scalebmp, Parameter.CELL_SIZE.Width, Parameter.CELL_SIZE.Height, Parameter.PART_NUMBER);
NormBlockVectorGram blockGram = new NormBlockVectorGram(hogGram, Parameter.BLOCK_SIZE.Width, Parameter.BLOCK_SIZE.Height);
DetectResultLayer detectlayer = new DetectResultLayer();
// !!!!!!检测窗口的像素尺寸必须能被cell尺寸整除!!!!!!像素尺寸除以hog尺寸就是检测窗口的尺寸
detectlayer.DetectResult = blockGram.DetectImgByHogWindow(
new Size(Parameter.DETECT_WINDOW_SIZE.Width / Parameter.CELL_SIZE.Width, Parameter.DETECT_WINDOW_SIZE.Height / Parameter.CELL_SIZE.Height),
NegCenter, PosCenter, Parameter.POS_DIST_COEF);
if (detectlayer.DetectResult == null)
{
return(null);
}
detectlayer.ScaleCoef = scalecoef;
resultlayers.Add(detectlayer); // 本层检测结果加入队列
scalecoef *= Parameter.SCALE_COEF; // 逐次缩小图像
newwidth = (int)(bmp.Width / scalecoef);
newheight = (int)(bmp.Height / scalecoef);
cnt++;
} while (newwidth > 2 * Parameter.DETECT_WINDOW_SIZE.Width && newheight > 2 * Parameter.DETECT_WINDOW_SIZE.Height);
elapse = DateTime.Now.Subtract(dt).TotalSeconds;
// 框出所有可能的物体
WindowResult[] wr = null;
Rectangle rect;
double mindist = -1;
WindowResult min_obj = null;
double min_scalecoef = 1;
resultCollection = new RectangleCollection();
foreach (DetectResultLayer layer in resultlayers)
{
wr = layer.DetectResult;
for (int i = 0; i < wr.Length; i++)
{
if (wr[i].label == 1)
{
if (mindist == -1 || mindist > wr[i].PosDistance)
{
mindist = wr[i].PosDistance;
min_obj = wr[i];
min_scalecoef = layer.ScaleCoef;
}
rect = new Rectangle((int)(wr[i].ImageRegion.X * layer.ScaleCoef * se),
(int)(wr[i].ImageRegion.Y * layer.ScaleCoef * se),
(int)(wr[i].ImageRegion.Width * layer.ScaleCoef * se),
(int)(wr[i].ImageRegion.Height * layer.ScaleCoef * se));
resultCollection.Add(rect);
}
}
}
//rect = new Rectangle((int)(min_obj.ImageRegion.X * min_scalecoef * se),
// (int)(min_obj.ImageRegion.Y * min_scalecoef * se),
// (int)(min_obj.ImageRegion.Width * min_scalecoef * se),
// (int)(min_obj.ImageRegion.Height * min_scalecoef * se));
//resultCollection.Add(rect);
return(resultCollection);
}
/// <summary>
/// 训练负样本
/// </summary>
/// <param name="bmp">负样本位图</param>
public void TrainNegative(Bitmap bmp)
{
Bitmap samplebmp = null;
double neg_distance = 0;
double pos_distance = 0;
bool hasinserted = false; // 指明样本是否已插入队列
if (bmp.Width / Parameter.DETECT_WINDOW_SIZE.Width > bmp.Height / Parameter.DETECT_WINDOW_SIZE.Height)
{
samplebmp = ImgOper.ResizeImage(bmp,
(int)(bmp.Width * Parameter.DETECT_WINDOW_SIZE.Height / bmp.Height), Parameter.DETECT_WINDOW_SIZE.Height);
}
else
{
samplebmp = ImgOper.ResizeImage(bmp, Parameter.DETECT_WINDOW_SIZE.Width,
(int)(bmp.Height * Parameter.DETECT_WINDOW_SIZE.Width / bmp.Width));
}
samplebmp = ImgOper.CutImage(samplebmp, 0, 0, Parameter.DETECT_WINDOW_SIZE.Width, Parameter.DETECT_WINDOW_SIZE.Height);
samplebmp = ImgOper.Grayscale(samplebmp);
HogGram hogGram = HogGram.GetHogFromBitmap(samplebmp, Parameter.CELL_SIZE.Width, Parameter.CELL_SIZE.Height, Parameter.PART_NUMBER);
NormBlockVectorGram blockGram = new NormBlockVectorGram(hogGram, Parameter.BLOCK_SIZE.Width, Parameter.BLOCK_SIZE.Height);
Rectangle rect = new Rectangle(0, 0, hogGram.HogSize.Width, hogGram.HogSize.Height);
double[] vect = blockGram.GetHogWindowVec(rect);
if (Dimension != 0 && vect.Length != Dimension)
{
throw new Exception("输入负样本的尺寸与其他样本尺寸不一致！");
}
ValuedBitmap vbmp = null;
if (PosCenter != null && NegCenter != null)
{
// 计算离正负中心的距离
for (int i = 0; i < vect.Length; i++)
{
neg_distance += Math.Abs(vect[i] - NegCenter[i]);
pos_distance += Math.Abs(vect[i] - PosCenter[i]);
}
// 与正样本中心重合时，说明是正样本，不能插入负样本队列
if (pos_distance == 0)
{
return;
}
// 负样本加入样本队列的第二道关，如果不够接近负样本中心，就无法加入队列
// 按照Hog检测的判定条件，正距离乘以Parameter.POS_DIST_COEF，使其避开边界
if (pos_distance * Parameter.POS_DIST_COEF < neg_distance)
{
return;
}
// 带归一化的系数，如果用neg_distance / pos_distance，值可能会溢出;
// 将neg_distance / (pos_distance + neg_distance)作为负样本的评价系数，值越小越接近负样本
vbmp = new ValuedBitmap(samplebmp, neg_distance / (pos_distance + neg_distance));
}
else
{
// 如果正样本库还没建立起来，则Val暂时赋值为1
vbmp = new ValuedBitmap(samplebmp, 1);
}
// 负样本加入样本队列的第三道关，与负样本评价系数的有序队列比较后，决定是否加入样本队列
hasinserted = InsertValuedBitmap(ref NegMapCollection, vbmp, Parameter.NEG_LIMITED_NUMBER);
NegLength = NegMapCollection.Count;
//// 人工观察负样本插入情况
//if (hasinserted && vbmp != null)
//{
// vbmp.VBitmap.Save("Image\\neg_save\\" + negcnt + "_" + vbmp.Val + ".jpg");
// negcnt++;
//}
// 如果样本已经插入队列，说明样本比较可信，重新计算样本中心
if (hasinserted)
{
if (NegCenter == null)
{
Dimension = vect.Length;
NegCenter = new double[Dimension];
}
for (int i = 0; i < Dimension; i++)
{
NegCenter[i] = (NegCenter[i] * NegLength + vect[i]) / (NegLength + 1);
}
}
}