Application of Deep Convolutional Neural Networks for automated and rapid identification and characterization of thin cracks in SHCCs

TL;DR

This paper demonstrates that tailored deep convolutional neural networks can rapidly and accurately identify and characterize surface cracks in SHCCs from photographs, improving upon manual and traditional image processing methods.

Contribution

The study develops and evaluates a tailored deep convolutional neural network for automatic crack detection and characterization in SHCCs, showing robustness and practical applicability.

Findings

High inference accuracy of TDCNN in crack detection

Robustness against epistemic uncertainty

Ability to compute crack width and density from images

Abstract

Previous research has showcased that the characterization of surface cracks is one of the key steps towards understanding the durability of strain hardening cementitious composites (SHCCs). Under laboratory conditions, surface crack statistics can be obtained from images of specimen surfaces through manual inspection or image processing techniques. Since these techniques require optimal lighting conditions, proper surface treatment, and prior (manual) selection of the correct region for proper inference, they are strenuous and time-consuming. Through this work, we explored and tailored deep convolutional networks (DCCNs) for the rapid characterization of cracks in SHCC from various kinds of photographs. The results from the controlled study suggest that the inference ability of the tailored DCCN (TDCNN) is quite good, resilient against epistemic uncertainty, and tunable for completely independent but adverse observations. From the crack pattern computed using TDCCN, average crack width (ACW) and crack density (CD) can be calculated to facilitate durability design and conditional assessment in a practical environment.