advanced image segmentation

[tischi]

Facts to be demonstrated

• saturation cannot be fixed. impossible to determine number of objects (and to quantify intensities)
• intensity offsets are normal; they usually are not a problem

Contents

• distance transform
  • example: nuclear periphery quantification
• macro recording of complex workflows

Formative assessments

• outcomes of convolution filters on some data
• assessing by etherpad

[tischi]

| 0 | 10 | 1 | | 10 | 11 | 10 | | 0 | 10 | 0 |

What is the outcome of a 3x3 mean filter for the central pixel of above tiny 8-bit image? a) 11 (wrong, simple the central pixel) b) 10 (wrong, this is the median and not the mean) c) 52 (wrong, this is the sum and not the mean) d) 8.66666 (wrong, did not divide sum by 9 but only by 6, which is the number of non-zero entries) e) 5.77777 (wrong-ish, this is the mean, but 8-bit images cannot have floating point values) f) 6 (correct)

[tischi]

Machine learning

• When it becomes too complicated (show many filters decision tree components map), consider using machine learning
• Show example (em)

[tischi]

Useful image filters

• invert
  • example: yeast cell wall for finding cells
• dilation (closing)
  • example: yeast cell wall for finding cells
• distance transform
  • ...