Issues in Evaluation of Crack Detection Methods and a Proposed Solution

Abstract

Automatic crack detection is an essential task for the effective maintenance of roads and structures. In recent years, deep learning has been widely applied to crack detection, and many models have been proposed. However, there isn't a standardized metric for the evaluation of crack detection methods nor reliable comparisons. This paper addresses this issue by examining the evaluation metrics suitable for cracks, selecting useful ones, and suggesting improvements. Additionally, a crack detection evaluation software was proposed and tested on the CRACK500 dataset, comparing FPHBN and DAUNet crack detection methods. The results confirmed the efficacy of the proposed metrics and the proposed software in helping researchers evaluate and compare different methods efficiently and effectively.