Pixel-Level Crack Detection and Quantification of Nuclear Containment with Deep Learning

Abstract

Crack detection based on deep learning is an advanced technology, and many scholars have proposed many methods for the segmentation of pavement cracks. However, due to the difference of image specifications and crack characteristics, some existing methods are not effective in detecting cracks of containment. To quickly detect cracks and accurately extract crack quantitative information, this paper proposes a crack detection model, called MA_CrackNet, based on deep learning and a crack quantitative analysis algorithm. MA_CrackNet is an end-to-end model based on multiscale fusions that achieve pixel-level segmentation of cracks. Experimental results show that the proposed MA_CrackNet has excellent performance in the crack detection task of nuclear containment, achieving a precision, recall, F1, and mean intersection-over-union (mIoU) of 86.07%, 89.96%, 87.97%, and 89.19%, respectively, outperforming other advanced semantic segmentation models. The quantification algorithm automatically measures the four characteristic indicators of the crack, namely, the length of the crack, the area, the maximum width, and the mean width and obtains reliable results.