Creat_local_deformable_model_xld.hdev例程相关学习

 index:.../Matching/Deformable/creat_local_deformable_model_xld.hdev
 建立一个可变形的模板并进行匹配
这个例程中一个子函数对MVTec做处理，这里面很有意思，代码如下。

Blue := [0,48,117]

Orange := [255,181,41]

* 

* Read the image containing the logo

read_image (Image, 'pen/pen-01')

get_image_size (Image, Width, Height)

* 

* Segment the logo

auto_threshold (Image, Regions, 10)

connection (Regions, Regions)

select_shape (Regions, SelectedRegions, ['row1','holes_num'], 'and', [80,0], [100,0])
用row1左上角的点和holes_num连接部件和孔的数量，给零就是互不连接的部分来选择 ，大概没验证代码。

sort_region (SelectedRegions, SelectedRegions, 'character', 'true', 'row')
排序 

closing_rectangle1 (SelectedRegions, Characters, 3, 3)

opening_rectangle1 (Characters, Characters, 5, 5)

* 

* Extract the contour

dilation_rectangle1 (Characters, RegionDilation, 5, 5)

erosion_rectangle1 (Characters, RegionErosion, 5, 5)

difference (RegionDilation, RegionErosion, RegionDifference)
膨胀收缩做差 

count_obj (RegionDifference, Number)

gen_empty_obj (LogoContoursTemp)

for Index := 1 to Number by 1

    select_obj (RegionDifference, ObjectSelected, Index)

    reduce_domain (Image, ObjectSelected, ImageReduced)

    edges_sub_pix (ImageReduced, Edges, 'canny', 1.5, 10, 20)
提取轮廓，提取出来结果是分离的edges 

    union_adjacent_contours_xld (Edges, UnionContours, 5, 1, 'attr_keep')
这里做了连接 

    close_contours_xld (UnionContours, ClosedContours)   
闭操作，这前后几句都是惯例 

    concat_obj (LogoContoursTemp, ClosedContours, LogoContoursTemp) 

endfor

* 

* Create the colored image of the logo

smallest_rectangle1_xld (LogoContoursTemp, Row1, Column1, Row2, Column2)
用上面结合出来的五个部件画了个包含他们的矩形
 

Row1 := min(Row1)

Column1 := min(Column1)

Row2 := max(Row2)

Column2 := max(Column2)

WidthLogoOrig := Column2 - Column1 + 1

ScaleFactor := real(WidthLogo) / WidthLogoOrig

hom_mat2d_identity (HomMat2DIdentity)

hom_mat2d_translate (HomMat2DIdentity, -0.5 * (Row1 + Row2), -0.5 * (Column1 + Column2), HomMat2DTranslate)

hom_mat2d_scale (HomMat2DTranslate, ScaleFactor, ScaleFactor, 0, 0, HomMat2DScale)

hom_mat2d_translate (HomMat2DScale, 0.5 * HeightImage, 0.5 * WidthImage, HomMat2DComplete)

affine_trans_contour_xld (LogoContoursTemp, LogoContours, HomMat2DComplete)

gen_image_const (ImageRed, 'byte', WidthImage, HeightImage)

gen_image_const (ImageGreen, 'byte', WidthImage, HeightImage)

gen_image_const (ImageBlue, 'byte', WidthImage, HeightImage)

compose3 (ImageRed, ImageGreen, ImageBlue, LogoImageTempl)

paint_xld (LogoContours, LogoImageTempl, LogoImage, [Blue,Orange,Blue,Blue,Blue,Blue])

* 

* Dilate the original logo

if (Broadening > 0)

    gray_dilation_shape (LogoImage, LogoImage, 2 * Broadening + 1, 2 * Broadening + 1, 'octagon')

elseif (Broadening < 0)

    gray_erosion_shape (LogoImage, LogoImage, -2 * Broadening + 1, -2 * Broadening + 1, 'octagon')

endif

return ()

学习了这个例程，程序很复杂了，就是建立一个可以变形的模板来做后期查找，但模板不好做

create_local_deformable_model_xld (LogoContours, 'auto', 0.0, 0.0, 'auto', 1, 1, 'auto', 1, 1, 'auto', 'auto', 'ignore_local_polarity', 5, [], [], ModelID)

find_local_deformable_model (GrayImage, ImageRectified, VectorField, DeformedContours, ModelID, 0, 0, 1, 1, 1, 1, 0.5, 1, 1, 4, 0.9, ['image_rectified','vector_field','deformed_contours'], ['deformation_smoothness','expand_border','subpixel'], [21,0,1], Score, Row, Column)

这两个函数是这个例程的核心，后面这个函数还做出来一个向量场 VectorField，grayimage是每次产生出来做对比的模板。
总之每次得分都很高，我个人感觉因为是商量MVTec都是源于一张图片，实际的话不知效果如何。