Fast concrete crack detection method via L2 sparse representation

A fast concrete crack detector based on L2 sparse representation is proposed. Specifically, via dividing the existing concrete images, many representative crack and non-crack image regions are selected for the over-complete dictionary. To suppress the noise disturbances, discrete cosine transformation is to extract the frequency-domain characteristics of these regions. For one new concrete image, it is first divided into many non-overlapping regions, and their sparse coefficients are fast computed on the established over-complete dictionary. Moreover then, a pooling operation is to extract the difference value between their sum coefficients on the crack templates and those on the non-crack ones, and easily yet effectively select the crack candidates via the sign bit of their difference values. Experiments on the practical concrete images show that the algorithm has high precision and efficiency.