Intelligent identification system of concrete cracks based on deep learning algorithm

To enhance the efficacy of crack detection techniques and tackle challenges such as selecting initial clustering centers in extraction algorithms and their sensitivity to varying image backgrounds, this research proposes an intelligent concrete crack recognition system utilizing deep learning methodologies. Among these, convolutional neural networks (CNNs) stand out for their representation learning capabilities, allowing them to classify input data according to a hierarchical structure. In a case study involving a bridge inspection project in a specific city, a comprehensive dataset of images was collected on-site. A CNN was then developed to create a classification model tailored for identifying concrete crack images, successfully classifying images captured in complex inspection environments. Furthermore, addressing issues related to high local cluster density and the Euclidean distance of clustering centers inherent in the traditional K-means algorithm, improvements were made by integrating statistical principles and morphological techniques. This enhanced K-means algorithm effectively facilitates the extraction of crack skeletons and the measurement of crack widths within challenging scene images. The method was verified by analyzing 600 photographs taken from the bridge's surface during the inspection, under complex conditions that included factors like surface peeling, stains, and moss—elements that typically complicate crack skeleton extraction. The results demonstrated a significant increase in crack detection efficiency compared to conventional methods.