Modular deep learning segmentation algorithm for concrete microscopic images

Segmentation procedures of concrete microscopic images and standard test methods devoted to the spacing factor calculation for freeze-thaw resistance assessment are time-consuming skill-dependent. Moreover, manual color treatment careful image examination often needed. Within past few years, Convolutional neural networks (CNN) have proved unpreceded performances in segmentation object detection tasks, though they showed limited reusability modularity. This study introduces an open-source modular deep learning algorithm images. The is based on two CNN models dedicated air voids aggregates detection. been calculated using various concrete, mortar, cement paste samples. Protected Paste Volume (PPV) distance-to-air-void computed agreed well with experimental factor. a better correlation between PPV scaling was found than experimentally measured factors scaling, highlighting critical interval from 200 µm 300 µm.