Surface concrete cracks detection and segmentation using transfer learning and multi-resolution image processing

Abstract

Surface crack detection must be precise and efficient for the structural health monitoring (SHM) of concrete structures. The traditional techniques of crack detection entail human visual examination. These procedures are labor-intensive, expensive, and inefficient. In order to forestall potentially hazardous circumstances caused by concrete surfaces that have deteriorated, it is vital to have an autonomous concrete crack inspection system. Crack detection and segmentation are the two main components of automatic crack inspection. In this research, deep convolutional neural networks (DCNN) and transfer learning (TL) methods are employed for crack detection in concrete structure images, also multiresolution image analysis based on the wavelet transform is the main process in the crack segmentation. The concrete crack dataset is randomly selected from two prevalent concrete crack datasets for developing the proposed method. The results indicated that DCNN classifier models provide good performance with F1-score ranging from 94.5 to 99.6 percent. In addition, the proposed multiresolution image analysis can segment crack pixels with 95.25 percent for the F1-score and shows acceptable and stable performance regarding various segmentation performance metrics.