请教一个 opencv 算法，图片中检测圆形对象


还请各位大佬不吝赐教~

你贴的什么，啥也看不到

YoLov5 应该可以，训练起来也简单

推荐 YoLov5 +1

霍夫圆的参数对结果影响很大，如果不需要拟合圆的话，直接 yolo 吧

yolov5 +1

如果全是这种有规律的图案，自己手搓代码（硬编码）速度应该可以在 0.1s 内。
你这个肉眼看特征很明显，就是白底上的小圆，二值化后找轮廓，然后手工写代码判断下是不是圆

从图片上看特征还是比较明显的，第一步可以行把方形小纸片扣出来，因为圆形的直径和小纸片长宽基本保持在一个比例范围，所以可以根据小纸片尺寸指定一个适合的霍夫算法参数，第二步再用霍夫检测圆，如果检测出多个圆形，第三步可以再次根据圆形和小纸片的面积或者尺寸比例 排除一下

intel 的 CPU 的话可以用 openvino 来加速，跑 onnx 模型没那么慢。用上面提到的 yolov5 ，应该不难。

精度要求不高的话可以考虑传统算法，用 opencv 进行模板匹配。
你可以先参考下： https://blog.csdn.net/yukinoai/article/details/88366234

我试了一下,霍夫曼检测好像没问题,就是不太准


代码是抄的 https://aistudio.baidu.com/projectdetail/649807

感谢大家~

打算先试下霍夫（图片先缩放一下），然后 15 楼的模版匹配，然后再试试 yolo5

大概率不用霍夫，这个准确度是在是难以达到。。。

这种程度直接模式匹配就行了，用不上机器学习都。opencv 自带的有模式匹配函数

Halcon 代码,可以试下

* This program fits circles into circular contour segments
* to extract their positions and radii
*
dev_update_off ()
*
* step: acquire image
*
read_image (Image, 'OXfE6Ph.png')

* auto_threshold (Image, all_Region, 15.0)
* connection (all_Region, all_ConnectedRegions)
* select_shape (all_ConnectedRegions, all_ConnectedRegions, 'area', 'and', 7000, 14000)
* count_obj (all_ConnectedRegions, all_NumberContours)
* for Index := 1 to all_NumberContours by 1
* select_obj(all_ConnectedRegions, ObjectSelected1, Index)
* smallest_rectangle1(ObjectSelected1, Row1, Column1, Row2, Column2)
* crop_part(Image, ImagePart, Row1, Column1, Row2-Row1, Column2-Colu)
* endfor

get_image_size (Image, Width, Height)
dev_close_window ()
dev_open_window (0, 0, Width / 2, Height / 2, 'light gray', WindowID)
dev_set_part (0, 0, Height - 1, Width - 1)
dev_set_line_width (3)
dev_set_color ('white')
dev_set_draw ('margin')
dev_display (Image)
set_display_font (WindowID, 14, 'mono', 'true', 'false')
disp_continue_message (WindowID, 'black', 'true')
stop ()
*
* step: segment image into regions
*
dev_set_colored (12)
dev_set_line_width (2)
dev_set_draw ('fill')

*equ_histo_image (Image, Image)
scale_image_max (Image, Image)
decompose3 (Image, Image_R, Image_G, Image_B)
* fast_threshold (Image, Region, 180, 230, 20)
auto_threshold (Image_G, Region, 5.0)
stop()

connection (Region, ConnectedRegions)
select_shape (ConnectedRegions, SelectedRegions, 'circularity', 'and', 0.6, 1.2)
select_shape (SelectedRegions, SelectedRegions, 'area', 'and', 200, 500)
dev_display (Image)
dev_display (SelectedRegions)
disp_continue_message (WindowID, 'black', 'true')
stop ()
*
* step: create ROI for contour processing
*
boundary (SelectedRegions, RegionBorder, 'inner_filled')
dilation_circle (RegionBorder, RegionDilation, 3.5)
union1 (RegionDilation, RegionUnion)
reduce_domain (Image, RegionUnion, ImageReduced)
dev_clear_window ()
dev_display (ImageReduced)
disp_continue_message (WindowID, 'black', 'true')
stop ()
*
* step: create contours and fit circles to them
*
edges_sub_pix (ImageReduced, Edges, 'canny', 1.5, 10, 40)
segment_contours_xld (Edges, ContoursSplit, 'lines_circles', 5, 2, 2)
select_contours_xld (ContoursSplit, SelectedContours, 'contour_length', 25, 99999, -0.5, 0.5)
count_obj (SelectedContours, NumberContours)
gen_empty_obj (Circles)
for i := 1 to NumberContours by 1
select_obj (SelectedContours, ObjectSelected, i)
get_contour_global_attrib_xld (ObjectSelected, 'cont_approx', Attrib)
if (Attrib == 1)
concat_obj (Circles, ObjectSelected, Circles)
endif
endfor
union_cocircular_contours_xld (Circles, UnionContours, rad(60), rad(10), rad(30), 100, 50, 10, 'true', 1)
dev_clear_window ()
dev_set_color ('black')
dev_display (UnionContours)
disp_continue_message (WindowID, 'black', 'true')
stop ()
count_obj (UnionContours, NumberCircles)
CenterRow := []
CenterColumn := []
dev_clear_window ()
dev_set_color ('black')
set_display_font (WindowID, 12, 'mono', 'true', 'false')
dev_display (SelectedContours)
for i := 1 to NumberCircles by 1
select_obj (UnionContours, ObjectSelected, i)
fit_circle_contour_xld (ObjectSelected, 'algebraic', -1, 0, 0, 3, 2, Row, Column, Radius, StartPhi, EndPhi, PointOrder)
gen_circle_contour_xld (ContCircle, Row, Column, Radius, 0, rad(360), 'positive', 1.5)
dev_set_color ('white')
dev_display (ContCircle)
if (i == 1)
Row2 := Row + Radius * sin(rad(-45))
Column2 := Column + Radius * cos(rad(-45))
set_tposition (WindowID, Row2 - 35, Column2 + 5)
endif
if (i > 1)
exist := 0
for j := 0 to i - 2 by 1
distance_pp (Row, Column, CenterRow[j], CenterColumn[j], DistanceCenters)
if (DistanceCenters < 20)
exist := 1
endif
endfor
if (exist == 1)
Row2 := Row + Radius * sin(rad(-135))
Column2 := Column + Radius * cos(rad(-135))
set_tposition (WindowID, Row2 - 40, Column2 - 30)
else
Row2 := Row + Radius * sin(rad(-45))
Column2 := Column + Radius * cos(rad(-45))
set_tposition (WindowID, Row2 - 35, Column2 + 5)
endif
endif
disp_arrow (WindowID, Row, Column, Row2, Column2, 2)
write_string (WindowID, i)
if (i < 8)
disp_message (WindowID, 'R' + i + ' = ' + Radius$'.4', 'window', i * 20, 130, 'black', 'false')
else
disp_message (WindowID, 'R' + i + ' = ' + Radius$'.4', 'window', (i - 7) * 20, 400, 'black', 'false')
endif
CenterRow := [CenterRow,Row]
CenterColumn := [CenterColumn,Column]
endfor
*
dev_update_window ('on')

你换个思路，在白色的区域检测是不是有较大色差，是不是比检测形状简单多了。

https://github.com/bubbliiiing/yolov7-pytorch 可以用这个训练，本地训练太慢可以在 https://www.autodl.com/home 上训练

可以看下 opencv 的二值化和轮廓

@Yuhyeong 谢谢
@sslzjx 谢谢，但是 halcon 不会，学习成本有点高了~
@lycpang 这个感觉比 yolo 还要难一点。。。
@czita 谢谢
@kyingstar 嗯。周末试下。谢谢