标签破损检测, 求处理思路

如图中标签粘贴时产生了破损，如何检测出这个破损区域，求大神点化，感谢


原图在这里:
https://pan.baidu.com/s/1NclWhaNPXQ-ms6SHaDbrBQ?pwd=1234 提取码: 1234

先定位对齐，再图像对比不行？参考https://www.51halcon.com/forum.php?mod=viewthread&tid=7960

1. read_image (Image, 'Image.jpg')
   
2. rgb3_to_gray (Image, Image, Image, ImageGray)
   
3. threshold (ImageGray, Region, 0, 200)
   
4. connection (Region, ConnectedRegions)
   
5. select_shape_std (ConnectedRegions, SelectedRegions, 'max_area', 70)
   
6. opening_circle (SelectedRegions, RegionOpening, 15)
   
7. closing_circle (RegionOpening, RegionClosing, 50)
   
8. reduce_domain (ImageGray, RegionClosing, ImageReduced)
   
9. threshold (ImageReduced, Region1, 200, 255)
   
10. connection (Region1, ConnectedRegions1)
    
11. select_shape (ConnectedRegions1, SelectedRegions1, 'area', 'and', 50, 10000)

复制代码

或者是存一个正常产品的区域，定位仿射再裁剪对比。

当前测试基本逻辑:
read_image (Image, '20260410-102844140_2569_f18_F2109_E13500-G0-P20260410_102844078_N1_BQJC_FCE25090832.png')

rgb1_to_gray (Image, GrayImage)

*根据ROI裁剪检测区域
gen_rectangle1 (ROI_0, 650.051, 1500.04, 3204.07, 3051.56)
reduce_domain (GrayImage, ROI_0, ImageReduced)

*最大暗区域即标签区域
fast_threshold (ImageReduced, Region, 0, 180, 200)
closing_rectangle1 (Region, RegionClosing, 1, 25)
fill_up (RegionClosing, RegionFillUp)
opening_circle (RegionFillUp, RegionOpening, 5.5)
closing_circle (RegionOpening, RegionClosing, 5.5)
connection (RegionClosing, ConnectedRegions)


*标签区域转凸包
select_shape_std (ConnectedRegions, SelectedRegions, 'max_area', 70)
shape_trans (SelectedRegions, LabelRegion, 'convex')
reduce_domain (ImageReduced, LabelRegion, ImageReduced1)

*取油墨区域
binary_threshold (ImageReduced1, Region1, 'max_separability', 'dark', UsedThreshold)
dilation_circle (Region1, RegionDilation, 1.5)

*无油墨区域
difference (SelectedRegions, RegionDilation, RegionDifference)
gray_features (RegionDifference, ImageReduced1, 'median', Value)

*填充油墨区
paint_region (RegionDilation, ImageReduced1, ImageResult, Value, 'fill')

*阈值分割
region_features (LabelRegion, 'height', Value)
mean_image (ImageResult, ImageMean, 3, 0.5*Value)
dyn_threshold (ImageResult, ImageMean, RegionDynThresh, 35, 'light')
opening_circle (RegionDynThresh, RegionOpening, 3.5)
closing_circle (RegionOpening, RegionClosing, 5.5)
connection (RegionClosing, ConnectedRegions)

*排除标签两端
SideShift := 8
region_features (LabelRegion, 'row1', Value1)
region_features (LabelRegion, 'row2', Value2)
region_features (LabelRegion, 'width', Value3)
select_shape (ConnectedRegions, SelectedRegions, ['row1','row2','width'], 'and', [Value1+SideShift,0,0.2*Value3], ['max',Value2-SideShift,'max'])


dev_display (Image)
dev_display (SelectedRegions)

halcon2016 发表于 2026-4-27 11:22
先定位对齐，再图像对比不行？参考https://www.51halcon.com/forum.php?mod=viewthread&tid=7960 ...

感谢，我研究下，图像中的三维码和文字区都是变化的，图像对比不好做吧

freekun139com 发表于 2026-4-30 08:58
当前测试基本逻辑:
read_image (Image, '20260410-102844140_2569_f18_F2109_E13500-G0-P20260410_102844 ...

你这种方法在实际应用中估计很难稳定吧，手动设置的参数太多了，你多传一些破损的图片看看，看有没有一些规律和共性，如果单看这一张，能想到的思路是用samsung做模板匹配，手动设置一个跟标签外框一样的roi做跟随区域，或者直接提取标签外框区域也行，最后在标签外框区域内直接斑点分析检测破损区域，你这图标签区域内，字是黑，背景是灰的，破损区域是白的，直接threshold就能提取破损区域

freekun139com 发表于 2026-4-30 09:01
感谢，我研究下，图像中的三维码和文字区都是变化的，图像对比不好做吧 ...

如果破损区域有可能发生在变化的码区和文字区，那对比确实不合适，如果不是，那可以用roi屏蔽掉变化的区域再对比

预定位到标签使用最小外接矩形，然后整体往里缩一个距离，检测内部有没有高亮的地方
行吗