//得到图片的HOG特征
public static Mat GetPlateSvmHOGFeatures(Mat img)
{
float[] descriptor = PlateCategory_SVM.ComputeHogDescriptors(img);
Mat feature = Float2Mat(descriptor);
return(feature);
}
//对是否为车牌进行训练
public static void TrainSVMDataForPlateRecog()
{
Console.WriteLine("preparing for training data!!!!");
List <TrainStruct> svmData = new List <TrainStruct>();
List <string> posImgFiles = getSampleFiles(@"E:\工作文件夹（workplace）\VSworkplace\PlateRecog\PlateRecog\普通车牌");
List <string> negImgFiles = getSampleFiles(@"E:\工作文件夹（workplace）\VSworkplace\PlateRecog\PlateRecog\非车牌");
//添加将正样本数据
for (int index = 0; index < posImgFiles.Count; index++)
{
TrainStruct trainData;
trainData.file = posImgFiles[index];
trainData.label = 1;
svmData.Add(trainData);
}
//添加负样本数据
for (int index = 0; index < negImgFiles.Count; index++)
{
TrainStruct trainData;
trainData.file = negImgFiles[index];
trainData.label = -1;
svmData.Add(trainData);
}
Mat samples = new Mat();
Mat responses = new Mat();
//读取数据并进行处理
for (int index = 0; index < svmData.Count; index++)
{
//以灰度的方式读取图片
Mat img = Cv2.ImRead(svmData[index].file, ImreadModes.Grayscale);
//剔除无用数据
if (img.Data == null)
{
Console.WriteLine("failed to load image {0}", svmData[index].file);
}
//对图片进行二值化
Mat dst = new Mat();
Cv2.Threshold(img, dst, 1, 255, ThresholdTypes.Otsu | ThresholdTypes.Binary);
//gray.Threshold(1, 255, ThresholdTypes.Otsu | ThresholdTypes.Binary)
Mat feature = GetPlateSvmHOGFeatures(dst);
//获取HOG特征
feature = feature.Reshape(1, 1);
samples.PushBack(feature);
responses.PushBack(Int2Mat(svmData[index].label));
}
// 训练数据的格式，OpenCV规定 samples 中的数据都是需要32位浮点型
// 因为TrainData::create 第一个参数是规定死的要cv_32F
samples.ConvertTo(samples, MatType.CV_32F);
// samples 将图片和样本标签合并成为一个训练集数据
// 第二个参数的原因是，我们的samples 中的每个图片数据的排列都是一行
if (PlateCategory_SVM.Train(samples, responses))
{
Console.WriteLine("Traing!!!");
PlateCategory_SVM.Save(@"E:\工作文件夹（workplace）\VSworkplace\PlateRecog\plateRecog.xml");
}
else
{
Console.WriteLine("failed to train data");
}
}
private static List <PlateInfo> LocatePlatesBySobel(Mat matSource,
int blur_Size = 5,
int sobel_Scale = 1,
int sobel_Delta = 0,
int sobel_X_Weight = 1,
int sobel_Y_Weight = 0,
int morph_Size_Width = 17,
int morph_Size_Height = 3,
int minWidth = 60, int maxWidth = 180,
int minHeight = 18, int maxHeight = 80,
float minRatio = 0.15f, float maxRatio = 0.70f)
{
List <PlateInfo> plateInfos = new List <PlateInfo>();
if (matSource.Empty())
{
return(plateInfos);
}
Mat blur = null;
Mat gray = null;
blur = matSource.GaussianBlur(new OpenCvSharp.Size(blur_Size, blur_Size), 0, 0,
BorderTypes.Default);
gray = blur.CvtColor(ColorConversionCodes.BGR2GRAY);
// 对图像进行Sobel 运算，得到的是图像的一阶水平方向导数。
// Generate grad_x and grad_y
MatType ddepth = MatType.CV_16S;
Mat grad_x = gray.Sobel(ddepth, 1, 0, 3, sobel_Scale, sobel_Delta, BorderTypes.Default);
Mat abs_grad_x = grad_x.ConvertScaleAbs();
Mat grad_y = gray.Sobel(ddepth, 0, 1, 3, sobel_Scale, sobel_Delta, BorderTypes.Default);
Mat abs_grad_y = grad_y.ConvertScaleAbs();
Mat grad = new Mat();
Cv2.AddWeighted(abs_grad_x, sobel_X_Weight, abs_grad_y, sobel_Y_Weight, 0, grad);
// 对图像进行二值化。将灰度图像（每个像素点有256 个取值可能）转化为二值图像（每个像素点仅有1 和0 两个取值可能）。
Mat threshold = grad.Threshold(0, 255, ThresholdTypes.Otsu | ThresholdTypes.Binary);
// 使用闭操作。对图像进行闭操作以后，可以看到车牌区域被连接成一个矩形装的区域。
Mat element = Cv2.GetStructuringElement(MorphShapes.Rect,
new OpenCvSharp.Size(morph_Size_Width,
morph_Size_Height));
Mat threshold_Close = threshold.MorphologyEx(MorphTypes.Close, element);
Mat element_Erode = Cv2.GetStructuringElement(MorphShapes.Rect, new OpenCvSharp.Size(5,
5));
Mat threshold_Erode = threshold_Close.Erode(element_Erode);
// Find 轮廓 of possibles plates 求轮廓。求出图中所有的轮廓。这个算法会把全图的轮廓都计算出来，因此要进行筛选。
OpenCvSharp.Point[][] contours = null;
HierarchyIndex[] hierarchys = null;
Cv2.FindContours(threshold_Erode, out contours, out hierarchys,
RetrievalModes.External, ContourApproximationModes.ApproxNone);
// 筛选。对轮廓求最小外接矩形，然后验证，不满足条件的淘汰。
for (int index = 0; index < contours.Length; index++)
{
Rect rectROI = Cv2.BoundingRect(contours[index]);
if (VerifyPlateSize(rectROI.Size,
minWidth,
maxWidth,
minHeight,
maxHeight,
minRatio,
maxRatio))
{
Mat matROI = matSource.SubMat(rectROI);
PlateCategory plateCategory = PlateCategory_SVM.Test(matROI);
if (plateCategory == PlateCategory.非车牌)
{
continue;
}
PlateInfo plateInfo = new PlateInfo();
plateInfo.OriginalRect = rectROI;
plateInfo.OriginalMat = matROI;
plateInfo.PlateCategory = plateCategory;
plateInfo.PlateLocateMethod = PlateLocateMethod.Sobel法;
plateInfos.Add(plateInfo);
}
}
return(plateInfos);
}
public static List <PlateInfo> LocatePlatesForAutoSample(Mat matSource,
out Mat matProcess,
int blur_Size = 5,
int sobel_Scale = 1,
int sobel_Delta = 0,
int sobel_X_Weight = 1,
int sobel_Y_Weight = 0,
int morph_Size_Width = 17,
int morph_Size_Height = 3,
int minWidth = 60, int maxWidth = 180,
int minHeight = 18, int maxHeight = 80,
float minRatio = 0.15f, float maxRatio = 0.70f)
{
List <PlateInfo> plateInfos = new List <PlateInfo>();
Mat gray = null;
Mat blur = null;
if (matSource.Empty() || matSource.Rows == 0 || matSource.Cols == 0)
{
matProcess = new Mat(0, 0, MatType.CV_8UC1);
return(plateInfos);
}
//灰度图
gray = matSource.CvtColor(ColorConversionCodes.BGR2GRAY);
//高斯模糊图
blur = gray.GaussianBlur(new OpenCvSharp.Size(blur_Size, blur_Size), 0, 0, BorderTypes.Default);
//HSV图
Mat hsv = matSource.CvtColor(ColorConversionCodes.BGR2HSV);
//将HSV图分成H、S、V通道图
Mat[] hsvSplits = hsv.Split();
hsvSplits[2] = hsvSplits[2].EqualizeHist();
Mat hsvEqualizeHist = new Mat();
//合并通道图
Cv2.Merge(hsvSplits, hsvEqualizeHist);
//提取蓝色
Scalar blueStart = new Scalar(100, 70, 70);
Scalar blueEnd = new Scalar(140, 255, 255);
Mat blue = hsvEqualizeHist.InRange(blueStart, blueEnd);
//提取黄色
Scalar yellowStart = new Scalar(15, 70, 70);
Scalar yellowEnd = new Scalar(40, 255, 255);
Mat yellow = hsvEqualizeHist.InRange(yellowStart, yellowEnd);
Mat add = blue + yellow;
//二值化
Mat threshold = add.Threshold(0, 255, ThresholdTypes.Otsu | ThresholdTypes.Binary);
Mat element = Cv2.GetStructuringElement(MorphShapes.Rect,
new OpenCvSharp.Size(morph_Size_Width, morph_Size_Height));
//闭操作
Mat threshold_Close = threshold.MorphologyEx(MorphTypes.Close, element);
Mat element_Erode = Cv2.GetStructuringElement(MorphShapes.Rect, new OpenCvSharp.Size(3,
3));
Mat threshold_Erode = threshold_Close.Erode(element_Erode);
matProcess = threshold_Erode;
OpenCvSharp.Point[][] contours = null;
HierarchyIndex[] hierarchys = null;
Cv2.FindContours(threshold_Erode, out contours, out hierarchys,
RetrievalModes.External, ContourApproximationModes.ApproxNone);
int isPlateCount = 0;
for (int index = 0; index < contours.Length; index++)
{
RotatedRect rotatedRect = Cv2.MinAreaRect(contours[index]);
Rect rectROI = Cv2.BoundingRect(contours[index]);
if (VerifyPlateSize(rectROI.Size,
minWidth,
maxWidth,
minHeight,
maxHeight,
minRatio,
maxRatio))
{
Mat matROI = matSource.SubMat(rectROI);
PlateCategory plateCategory = PlateCategory_SVM.Test(matROI);
if (plateCategory != PlateCategory.非车牌)
{
isPlateCount++;
}
PlateInfo plateInfo = new PlateInfo();
plateInfo.RotatedRect = rotatedRect;
plateInfo.OriginalRect = rectROI;
plateInfo.OriginalMat = matROI;
plateInfo.PlateCategory = plateCategory;
plateInfo.PlateLocateMethod = PlateLocateMethod.颜色法;
plateInfos.Add(plateInfo);
}
}
if (isPlateCount > 0)
{
return(plateInfos);
}
blur = matSource.GaussianBlur(new OpenCvSharp.Size(blur_Size, blur_Size), 0, 0,
BorderTypes.Default);
gray = blur.CvtColor(ColorConversionCodes.BGR2GRAY);
// 对图像进行Sobel 运算，得到的是图像的一阶水平方向导数。
// Generate grad_x and grad_y
MatType ddepth = MatType.CV_16S;
Mat grad_x = gray.Sobel(ddepth, 1, 0, 3, sobel_Scale, sobel_Delta, BorderTypes.Default);
Mat abs_grad_x = grad_x.ConvertScaleAbs();
Mat grad_y = gray.Sobel(ddepth, 0, 1, 3, sobel_Scale, sobel_Delta, BorderTypes.Default);
Mat abs_grad_y = grad_y.ConvertScaleAbs();
Mat grad = new Mat();
Cv2.AddWeighted(abs_grad_x, sobel_X_Weight, abs_grad_y, sobel_Y_Weight, 0, grad);
// 对图像进行二值化。将灰度图像（每个像素点有256 个取值可能）转化为二值图像（每个像素点仅有1 和0 两个取值可能）。
threshold = grad.Threshold(0, 255, ThresholdTypes.Otsu | ThresholdTypes.Binary);
// 使用闭操作。对图像进?行行闭操作以后，可以看到车牌区域被连接成一个矩形装的区域。
element = Cv2.GetStructuringElement(MorphShapes.Rect,
new OpenCvSharp.Size(morph_Size_Width,
morph_Size_Height));
threshold_Close = threshold.MorphologyEx(MorphTypes.Close, element);
element_Erode = Cv2.GetStructuringElement(MorphShapes.Rect, new OpenCvSharp.Size(5, 5));
threshold_Erode = threshold_Close.Erode(element_Erode);
matProcess = threshold_Erode;
// Find 轮廓 of possibles plates 求轮廓。求出图中所有的轮廓。这个算法会把全图的轮廓都计算出来，因此要进? 行行筛选。
contours = null;
hierarchys = null;
Cv2.FindContours(threshold_Erode, out contours, out hierarchys,
RetrievalModes.External, ContourApproximationModes.ApproxNone);
// 筛选。对轮廓求最?小外接矩形，然后验证，不不满?足条件的淘汰。
for (int index = 0; index < contours.Length; index++)
{
Rect rectROI = Cv2.BoundingRect(contours[index]);
if (VerifyPlateSize(rectROI.Size,
minWidth,
maxWidth,
minHeight,
maxHeight,
minRatio,
maxRatio))
{
RotatedRect rotatedRect = Cv2.MinAreaRect(contours[index]);
Mat matROI = matSource.SubMat(rectROI);
PlateCategory plateCategory = PlateCategory_SVM.Test(matROI);
PlateInfo plateInfo = new PlateInfo();
plateInfo.RotatedRect = rotatedRect;
plateInfo.OriginalRect = rectROI;
plateInfo.OriginalMat = matROI;
plateInfo.PlateCategory = plateCategory;
plateInfo.PlateLocateMethod = PlateLocateMethod.Sobel法;
plateInfos.Add(plateInfo);
}
}
return(plateInfos);
}
//颜色法
private static List <PlateInfo> LocatePlatesByColor(Mat matSource,
int blur_Size = 5,
int morph_Size_Width = 17,
int morph_Size_Height = 3,
int minWidth = 60, int maxWidth = 180,
int minHeight = 18, int maxHeight = 80,
float minRatio = 0.15f, float maxRatio = 0.70f)
{
List <PlateInfo> plateInfos = new List <PlateInfo>();
if (matSource.Empty())
{
return(plateInfos);
}
Mat hsv = matSource.CvtColor(ColorConversionCodes.BGR2HSV);
Mat[] hsvSplits = hsv.Split();
hsvSplits[2] = hsvSplits[2].EqualizeHist();
Mat hsvEqualizeHist = new Mat();
Cv2.Merge(hsvSplits, hsvEqualizeHist);
Scalar blueStart = new Scalar(100, 70, 70);
Scalar blueEnd = new Scalar(140, 255, 255);
Mat blue = hsvEqualizeHist.InRange(blueStart, blueEnd);
Scalar yellowStart = new Scalar(15, 70, 70);
Scalar yellowEnd = new Scalar(40, 255, 255);
Mat yellow = hsvEqualizeHist.InRange(yellowStart, yellowEnd);
Mat add = blue + yellow;
Mat threshold = add.Threshold(0, 255, ThresholdTypes.Otsu | ThresholdTypes.Binary);
Mat element = Cv2.GetStructuringElement(MorphShapes.Rect,
new OpenCvSharp.Size(morph_Size_Width,
morph_Size_Height));
Mat threshold_Close = threshold.MorphologyEx(MorphTypes.Close, element);
Mat element_Erode = Cv2.GetStructuringElement(MorphShapes.Rect, new OpenCvSharp.Size(3,
3));
Mat threshold_Erode = threshold_Close.Erode(element_Erode);
OpenCvSharp.Point[][] contours = null;
HierarchyIndex[] hierarchys = null;
Cv2.FindContours(threshold_Erode, out contours, out hierarchys,
RetrievalModes.External, ContourApproximationModes.ApproxNone);
for (int index = 0; index < contours.Length; index++)
{
Rect rectROI = Cv2.BoundingRect(contours[index]);
if (VerifyPlateSize(rectROI.Size,
minWidth,
maxWidth,
minHeight,
maxHeight,
minRatio,
maxRatio))
{
Mat matROI = matSource.SubMat(rectROI);
PlateCategory plateCategory = PlateCategory_SVM.Test(matROI);
if (plateCategory == PlateCategory.非车牌)
{
continue;
}
PlateInfo plateInfo = new PlateInfo();
plateInfo.OriginalRect = rectROI;
plateInfo.OriginalMat = matROI;
plateInfo.PlateCategory = plateCategory;
plateInfo.PlateLocateMethod = PlateLocateMethod.颜色法;
plateInfos.Add(plateInfo);
}
}
return(plateInfos);
}