Crack Segmentation

A deep learning model for segmenting cracks. This repository will have more models in future.

Demo

• Full demo: https://bit.ly/choiw-fulldemo

• Crack-propagation1:
  

• Crack-propagation2:
  

• Crack-moving shot1:
  

• Crack-moving shot2:
  

• Asphalt-moving shot:
  

Libraries

• Python v3.x
• TensorFlow v1.15
• TensorFlow-Probability 0.7.0
• Horovod (https://github.com/horovod/horovod)
• OpenCV
• NumPy
• SciPy

Key techniques

• Half-precision (FP16)
• Feature pyramid
• Global context block
• Bottleneck layer
• Shortcut connection (concatenation and addition)

Model summary

• Dataset: built from scratch for this side project
• Architecture: depth-42 (38 downscale and 4 upscale), no pretrained model used
• Number of parameters: 2 million