Crack detection for nuclear containments based on multi-feature fused semantic segmentation

Crack detection on the outer surface of a nuclear containment to ensure nuclear power plant safety is an important task. However, as cracks on the surface of nuclear containments are relatively small, surrounding noise is substantial and contrast is low, traditional detection methods based on filters are complex, and their result are inadequate. Semantic segmentation, which is widely used in deep learning, can achieve fast and accurate crack segmentation at the pixel level. However, most of the existing networks are applied to pavement cracks, for the nuclear containment cracks with special characteristics, there is still much room for improvement. To improve crack detection performance for nuclear containments, this study uses close-range photogrammetry to obtain crack images and generate a crack dataset (3,240 images). Then, the comparison experiment proves that when the depth of the model reaches a level, the segmentation performance of the model does not increase with the increase of the depth. Therefore, this paper proposes a semantic segmentation model based on multi-feature fusion and focal loss. Through comparative experiments with several other commonly used methods and generalization experiments using other crack dataset, it is proved that the method proposed in this paper has better crack segmentation performance.

Automated summary

This study focuses on crack detection on the outer surface of a nuclear containment in order to ensure nuclear power plant safety. A semantic segmentation model based on multi-feature fusion and focal loss is proposed to improve the crack segmentation performance. Comparative experiments and generalization experiments prove that the proposed method performs better than other commonly used methods.