A robust self-driven surface crack detection algorithm using local features

This paper presents an effective image analysis method for visual surface crack detection, called a robust self-driven crack detection algorithm (RSCDA). Firstly, a local texture anisotropy (LTA) is estimated based on self-driven local feature statistics from the original photograph. Secondly, the LTA is used to detect candidate crack pixels. Finally, the actual crack pixels are accurately identified using two effective measurements for connected domains based on discriminative direction and relative sparse features. The results demonstrate that the RSCDA is an effective and robust surface crack detection method for building materials or textiles.