crack_detection

a ML based pavement crack detection

data prepration

• Positive samples: patch with black pixels (crack) at center

• Negative samples: patch with white pixels (non-crack) at center

• CNN: multi-label classification

  • Input : SxS patch around a sample (shape: SxSx3)

  • Output: structure around the sample in a KxK patch (shape: K^2 )

• HOG: binary classification

  • Input : HOG features from SxS patch around a sample

  • Output: class label of the sample (crack: c=0, non-crack: c=1)

• data collection

Testing on an image

• Use each pixel to generate an input patch

• Predict output value for each patch

• For HOG:

  • if output > 0.5 ==> pixel is crack
  • if output < 0.5 ==> pixel is non-crack

• For CNN:

  • multiple output values (overlapping)
  • at each pixel ouput value is affected from KxK surrounding pixels
  • summation and normalization ==> probability map
  • if pmap > 0.5 ==> pixel is crack
  • if pmap < 0.5 ==> pixel is non-crack

• Results:

• CNN (S=5)

• HOG (S=1)

• HOG (S=5)