求助，L形的不规则图形如何进行动态阈值

L形区域抠图后用动态阈值查找是否有亮斑，拐角处会被筛选出来，如何处理这种场景
这种场景是不是做平滑的时候受domain外区域影响了，求助大佬帮忙解释一下原理
这里如果把动态阈值调整到15就可以规避，想学习一下原理，以便遇到类似问题

主要代码如下：

1. read_image (Image, 'L形图像动态阈值.png')
   
2. decompose3 (Image, ImageR, ImageG, ImageB)
   
3. 
   
4. binary_threshold (ImageB, RegionBinThresh, 'max_separability', 'light', UsedThreshold)
   
5. opening_rectangle1 (RegionBinThresh, RegionOpening, 1, 15)
   
6. opening_rectangle1 (RegionOpening, RegionOpening1, 15, 1)
   
7. connection (RegionOpening1, ConnectedRegions2)
   
8. select_shape (ConnectedRegions2, SelectedRegions2, 'area', 'and', 5000, 999999)
   
9. 
   
10. union1 (SelectedRegions2, RegionUnion)
    
11. smallest_rectangle2 (RegionUnion, Row, Column, Phi, Length1, Length2)
    
12. gen_rectangle2 (Rectangle, Row, Column, Phi, Length1, Length2-8)
    
13. intersection (SelectedRegions2, Rectangle, RegionIntersection)
    
14. 
    
15. sort_region (RegionIntersection, SortedRegions, 'first_point', 'true', 'column')
    
16. region_features (SortedRegions, 'width', Value1)
    
17. tuple_mean (Value1, FinWidth)
    
18. KernelHeight := int(FinWidth) / 2 * 2 + 1
    
19. 
    
20. count_obj (SortedRegions, Number)
    
21. gen_empty_obj (DefectRegions)
    
22. for Index1 := 1 to Number by 1
    
23.     select_obj (SortedRegions, ObjectSelected, Index1)
    
24.    
    
25.     reduce_domain (ImageB, ObjectSelected, ImageReduced)
    
26.     crop_domain (ImageReduced, ImagePart)
    
27.     region_features (ObjectSelected, 'row1', Row1)
    
28.     region_features (ObjectSelected, 'column1', Column1)
    
29. 
    
30.     mean_image (ImagePart, ImageMean, 3, 3)
    
31.     mean_image (ImagePart, ImageMean1, 3, KernelHeight)
    
32.     dyn_threshold (ImageMean, ImageMean1, RegionDynThresh, 12, 'light')
    
33.     opening_circle (RegionDynThresh, RegionOpening2, 1.5)
    
34.     closing_circle (RegionOpening2, RegionClosing2, 2.5)
    
35.     connection (RegionClosing2, ConnectedRegions1)
    
36.     fill_up (ConnectedRegions1, RegionFillUp)
    
37.    
    
38.     select_shape (RegionFillUp, SelectedRegions1, 'area', 'and', 100, 99999)
    
39.     count_obj (SelectedRegions1, Number1)
    
40.     if (Number1 > 0)
    
41.         move_region (SelectedRegions1, SelectedRegions1, Row1, Column1)
    
42.         concat_obj (DefectRegions, SelectedRegions1, DefectRegions)
    
43.     endif
    
44.    
    
45. endfor
    
46. 
    
47. dev_display (Image)
    
48. dev_display (DefectRegions)

复制代码

筛选效果图.png

应该是这样的

吐个泡泡 发表于 2023-5-5 20:52
应该是这样的

感谢大佬指点，谢谢！
动态阈值原理是清楚的，这里不明白的是做平滑的时候对于不规则的图形，domain外的区域是怎么影响的(测试中crop后的图像domain也应该是L形的吧)
这里我理解是平滑算子的核高度大于了那块L形位置的高度，导致domain外的区域影响平滑后的图像，即这个平滑算子不保边

求老师发一个最新的DeepLearningTool23.04

官网去下载不就好了：
https://www.mvtec.com/cn/downloads/deep-learning-tool

学习一下