HALCON示例程序classify_citrus_fruits.hdev应用常规gmm分类器进行水果分类

示例程序源码（加注释）

• 读入图片、显示与显示相关设置
  read_image (Image, ‘color/citrus_fruits_01’)
  get_image_pointer1 (Image, Pointer, Type, Width, Height)
  dev_close_window ()
  dev_open_window (0, 0, Width, Height, ‘white’, WindowHandle)
  set_display_font (WindowHandle, 12, ‘mono’, ‘true’, ‘false’)
  dev_set_draw (‘margin’)
  dev_set_line_width (2)
  dev_display (Image)
  dev_update_window (‘off’)
  dev_update_pc (‘off’)
  dev_update_var (‘off’)

• 定义变量数组
  FeaturesArea := []
  FeaturesCircularity := []
  ClassName := [‘orange’,‘lemon’]

• 创建一个GMM分类器，这个函数之前的帖子介绍过
  create_class_gmm (2, 2, 1, ‘spherical’, ‘normalization’, 10, 42, GMMHandle)

• 为分类器增加例子
  for i := 1 to 4 by 1
  read_image (Image, ‘color/citrus_fruits_’ + i$’.2d’)
  dev_display (Image)

  • 自定义函数
    get_regions (Image, SelectedRegions)
    将彩色图像分割成RGB三个通道，并使用简单bolo方法进行分析
    decompose3 (Image, ImageRed, ImageGreen, ImageBlue)
    dev_set_color (‘white’)
    threshold (ImageRed, Region, 50, 255)
    fill_up (Region, RegionFillUp)
    connection (RegionFillUp, ConnectedRegions)
    select_shape (ConnectedRegions, SelectedRegions, ‘area’, ‘and’, 50, 999999)
    return ()
  • 显示SelectedRegions区域
    dev_display (SelectedRegions)
  • 对区域SelectedRegions进行计数
    count_obj (SelectedRegions, NumberObjects)
    for j := 1 to NumberObjects by 1
    • 选取数组当中的指定元素
      select_obj (SelectedRegions, ObjectSelected, j)
    • 自定义函数
      get_features (ObjectSelected, WindowHandle, Circularity, Area, RowRegionCenter, ColumnRegionCenter)
      得到ObjectSelected区域的圆度值
      circularity (ObjectSelected, Circularity)
      求取ObjectSelected的面积与中心坐标
      area_center (ObjectSelected, Area, Row, Column)
      dev_set_color (‘white’)
      return ()
      FeaturesArea := [FeaturesArea,Area]
      FeaturesCircularity := [FeaturesCircularity,Circularity]
    • real将输入数组元素转换为浮点数形式
      FeatureVector := real([Circularity,Area])
      if (i <= 2)
      • add_sample_class_gmm - 将训练样本添加到高斯混合模型的训练数据中
      • add_sample_class_gmm（:: GMM句柄，样本要素向量，类别ID，随机噪声：）
        add_sample_class_gmm (GMMHandle, FeatureVector, 0, 0)
      • 显示信息
        disp_message (WindowHandle, ‘Add to Class:’ + ClassName[0], ‘window’, RowRegionCenter, ColumnRegionCenter - 100, ‘black’, ‘true’)
        else
        add_sample_class_gmm (GMMHandle, FeatureVector, 1, 0)
        disp_message (WindowHandle, ‘Add to Class:’ + ClassName[1], ‘window’, RowRegionCenter, ColumnRegionCenter - 100, ‘black’, ‘true’)
        endif
        endfor
  • 显示按F5继续
    disp_continue_message (WindowHandle, ‘black’, ‘true’)
    stop ()
    endfor
    dev_clear_window ()

• 构建可视化空间，即二维坐标系
  visualize_2D_feature_space (Cross, Height, Width, WindowHandle, FeaturesArea[0:5], FeaturesCircularity[0:5], ‘dim gray’, 18)
  dev_set_color (‘black’)
  OriginOfGraph := [Height - 0.1 * Height,0.1 * Width]
  画箭头
  disp_arrow (WindowID, OriginOfGraph[0], OriginOfGraph[1], OriginOfGraph[0], Width - 0.2 * Width, 2)
  disp_arrow (WindowID, OriginOfGraph[0], OriginOfGraph[1], 0.1 * Height, OriginOfGraph[1], 2)
  显示横纵坐标的名称
  set_tposition (WindowID, OriginOfGraph[0], Width - 0.2 * Width)
  write_string (WindowID, ‘Area’)
  set_tposition (WindowID, 0.07 * Height, OriginOfGraph[1])
  write_string (WindowID, ‘Circularity’)
  dev_set_color (ColorFeatureVector)
  ExtentOfGraph := Height - 0.3 * Height
  RangeC := 0.5
  RangeA := 24000
  ScaleC := ExtentOfGraph / RangeC
  ScaleA := ExtentOfGraph / RangeA
  MinC := 0.5
  MinA := 20000
  NumberFeatureVectors := |FeaturesA|
  for i := 0 to NumberFeatureVectors - 1 by 1
  DiffC := ScaleC * (FeaturesC[i] - MinC)
  DiffA := ScaleA * (FeaturesA[i] - MinA)
  RowFeature := OriginOfGraph[0] - DiffC
  ColumnFeature := OriginOfGraph[1] + DiffA
  画X号
  gen_cross_contour_xld (Cross, RowFeature, ColumnFeature, CrossSize, 0.785398)
  dev_display (Cross)
  endfor
  return ()

• 柠檬的自定义函数
  visualize_2D_feature_space (Cross, Height, Width, WindowHandle, FeaturesArea[6:11], FeaturesCircularity[6:11], ‘light gray’, 18)
  disp_continue_message (WindowHandle, ‘black’, ‘true’)
  stop ()

• 训练分类器train_class_gmm（:: 分类器句柄，最大迭代次数，阈值，计算方法，规范单一协方差矩阵：类中心数量，每个类的迭代次数）
  train_class_gmm (GMMHandle, 100, 0.001, ‘training’, 0.0001, Centers, Iter)

• 分类
  for i := 1 to 15 by 1
  read_image (Image, ‘color/citrus_fruits_’ + i$’.2d’)
  dev_display (Image)

  • 和上文get_regions含义一致
    get_regions (Image, SelectedRegions)
    dev_display (SelectedRegions)
    count_obj (SelectedRegions, NumberObjects)
    for j := 1 to NumberObjects by 1
    select_obj (SelectedRegions, ObjectSelected, j)
    get_features (ObjectSelected, WindowHandle, Circularity, Area, RowRegionCenter, ColumnRegionCenter)
    FeaturesArea := [FeaturesArea,Area]
    FeaturesCircularity := [FeaturesCircularity,Circularity]
    FeatureVector := real([Circularity,Area])
    • classify_class_gmm - 通过高斯混合模型计算特征向量的类。
    • 函数原型classify_class_gmm（:: 分类器句柄，特征向量，确定的最佳类数：分类结果，类的后验概率，概率密度，归一化概率）
      classify_class_gmm (GMMHandle, FeatureVector, 1, ClassID, ClassProb, Density, KSigmaProb)
    • 显示
      disp_message (WindowHandle, 'Class: ’ + ClassName[ClassID], ‘window’, RowRegionCenter, ColumnRegionCenter - 100, ‘black’, ‘true’)
      disp_message (WindowHandle, 'KSigmaProb: ’ + KSigmaProb, ‘window’, RowRegionCenter + 30, ColumnRegionCenter - 100, ‘black’, ‘true’)
      endfor
      if (i != 15)
      disp_continue_message (WindowHandle, ‘black’, ‘true’)
      endif
      stop ()
      endfor

• 清除训练句柄
  clear_class_gmm (GMMHandle)

处理思路

这个例子介绍了GMM分类器的使用方法，步骤是create_class_gmm 、add_sample_class_gmm、train_class_gmm、classify_class_gmm 先创建分类器，为分类器添加样本，训练分类器，使用分类器对进行分类。一个全面介绍GMM分类器使用的一个例子。通过GMM分类器实现了橙子和柠檬 的分类。

后记

大家有什么问题可以向我提问哈，我看到了第一时间回复，希望在学习的路上多多结交良师益友。