Concrete crack analysis using a deep belief convolutional neural network

Abstract

The assessment of surface cracks in concrete structures plays a pivotal role in determining structural integrity. However, current diagnostic technologies suffer from drawbacks such as being time-consuming, subjective, and reliant on inspectors' experience, resulting in low detection accuracy. This paper seeks to address these issues by proposing an automated, vision-based method for identifying the surface condition of concrete structures. The method integrates advanced pre-trained convolutional neural networks (CNNs), transfer learning, and decision-level image fusion. To develop and validate this approach, a total of 6,500 image patches from diverse concrete surfaces were generated. Each pre-trained CNN establishes a predictive model for the initial diagnosis of surface conditions through transfer learning. Given the potential for conflicting results among different CNNs due to architectural differences, a modified Deep Belief CNN algorithm is crafted, thereby enhancing crack detection accuracy. The effectiveness of the proposed method is confirmed through a comparison with other CNN models. Robustness is tested by subjecting the method to images with various types and intensities of noise, yielding satisfactory outcomes. In practical scenarios, the hybridised approach is applied to analyse field-captured images of concrete structures using an exhaustive search-based scanning window. Results showcase the method's capacity to accurately identify crack profiles, with minimal areas of incorrect predictions underscoring its potential for practical applications.