Vision-based concrete crack detection using a hybrid framework considering noise effect

Diagnosing surface cracks of concrete structures has been a critical aspect of assessing structural integrity. Existing diagnosis technologies are time-consuming, subjective, and heavily dependent on the experiences of inspectors, which leads to low detection accuracy. This paper aims to resolve these challenges by proposing a vision-based automated method for surface condition identification of concrete structures, consisting of state-of-the-art pre-trained convolutional neural networks (CNNs), transfer learning, and decision-level image fusion. For this purpose, a total of 41,780 image patches of various concrete surfaces are generated for the development and validation of the proposed method. Each pre-trained CNN is employed to establish the predictive model for the initial diagnosis of surface conditions via transfer learning. Since different CNNs may generate conflicting results due to differences in network architectures, a modified Dempster-Shafer (D-S) algorithm is designed to conduct decision-level image fusion to improve the crack detection accuracy. The superiority of the proposed method is validated via the com-parison with single CNN models. The robustness of the proposed method is also verified using the images polluted with various types and intensities of noise, with satisfactory outcomes. Finally, this hybridised approach is applied to the analysis of images of concrete structures captured in the field, through an exhaust search-based scanning window. The results show that it is capable of accurately identifying the crack profile with wrong predictions of limited areas, demonstrating its potential in practical applications.

Automated summary

This paper proposes a vision-based automated method for surface condition identification of concrete structures, which utilizes pre-trained CNNs, transfer learning, and decision-level image fusion to improve the accuracy of crack detection. The method is validated using a large dataset and shows superiority over single CNN models.