Halcon knowledge: NCC_ Model template matching

  explain

         In addition to common template matching , There are other templates matching , Although not mainstream , But try to use , It can also solve many problems . Let's start with ncc_model Templates .

One 、 Based on cross-correlation matching

         A method based on gray value is based on correlation matching . This method uses normalized cross-correlation , Used to evaluate the correspondence between the model and the search image . It is significantly faster than the classic gray value based matching , And it can compensate for the multiplicative changes of addition and illumination . Contrary to shape based matching , Objects also have slightly changed shapes 、 Objects in a large number of textured or blurred images （ The outline disappears in the blur image , for example , Because of defocus ） Can find . applications ： The search object has slight deformation , Lots of textures , Image blur and other occasions , Fast , Low precision . Basic operation ：

1. Create cross-correlation templates ：create_ncc_model()

2. Look for correlations ：find_ncc_model()

3. Release template ：clear_ncc_model()

         In this section , We will quickly outline the matching process based on correlation matching . To actively follow the example , start-up HDevelop Program hdevelop\Matching\Correlation-Based\ find_ncc_model_defocused.hdev, It shows the robustness against texture and defocus based on correlation matching .

dev_update_off ()
read_image (Image, 'smd/smd_on_chip_05')
get_image_size (Image, Width, Height)
dev_close_window ()
dev_open_window (0, 0, Width, Height, 'black', WindowHandle)
set_display_font (WindowHandle, 16, 'mono', 'true', 'false')
dev_set_color ('green')
dev_set_draw ('margin')
gen_rectangle1 (Rectangle, 175, 156, 440, 460)
area_center (Rectangle, Area, RowRef, ColumnRef)
reduce_domain (Image, Rectangle, ImageReduced)
create_ncc_model (ImageReduced, 'auto', 0, 0, 'auto', 'use_polarity', ModelID)
dev_display (Image)
dev_display (Rectangle)
disp_continue_message (WindowHandle, 'black', 'true')
stop ()
for J := 1 to 11 by 1
    read_image (Image, 'smd/smd_on_chip_' + J$'02')
    find_ncc_model (Image, ModelID, 0, 0, 0.5, 1, 0.5, 'true', 0, Row, Column, Angle, Score)
    vector_angle_to_rigid (RowRef, ColumnRef, 0, Row, Column, 0, HomMat2D)
    affine_trans_region (Rectangle, RegionAffineTrans, HomMat2D, 'nearest_neighbor')
    dev_display (Image)
    dev_display (RegionAffineTrans)
    if (J < 11)
        disp_continue_message (WindowHandle, 'black', 'true')
    endif
    stop ()
endfor
clear_ncc_model (ModelID)

Two 、 Brief description of the procedure

step 1： Select the object in the reference image

         First , In the training image , Use gen_rectangle1 Create an area containing objects . This use area_center Query the center of the area . In the following steps, you will need to cover the matching between the result and the original area . then , Reduce the image to the area of interest .

read_image (Image, 'smd/smd_on_chip_05')

gen_rectangle1 (Rectangle, 175, 156, 440, 460)

area_center (Rectangle, Area, RowRef, ColumnRef)

reduce_domain (Image, Rectangle, ImageReduced)

step 2： Creating models

The reduced image is used to pass create_ncc_model establish NCC Model . result , Operator returns the newly created model （ModelID） The handle of , It can then be used to specify the model , for example , When calling the operator find_ncc_model when .

 create_ncc_model (ImageReduced, 'auto', 0, 0, 'auto', 'use_polarity', ModelID)

step 3： Find the object again

         Now? , Cycle through the image , And search based on the above template , For each search image , Search for NCC Model and use affine transformation to cover the model area . Please note that , Training is applied to focused images , The search is applied to defocus of image changes with the following characteristics . For all that , Object instances have been found . chart 3.1 It shows the reference image and one of the model examples that are defocused but found .

for J := 1 to 11 by 1

read_image (Image, 'smd/smd_on_chip_' + J$'02')

find_ncc_model (Image, ModelID, 0, 0, 0.5, 1, 0.5, 'true', 0, Row, \

Column, Angle, Score)

vector_angle_to_rigid (RowRef, ColumnRef, 0, Row, Column, 0, HomMat2D)

affine_trans_region (Rectangle, RegionAffineTrans, HomMat2D, \

'nearest_neighbor')

dev_display (Image)

dev_display (RegionAffineTrans)

endfor

 

step 4： Destroy the model

When there is no need for NCC Model time , Use clear_ncc_model Destroy it .clear_ncc_model (ModelID)