Enhancing Dam Inspection with Pixel-Level CNN-FCN Approach Via 3D Texture Mapping

Abstract

Inspecting damage through visual inspection is an outdated and inefficient method, especially for large concrete dam structures. Furthermore, inspecting for damage to dam structures is costly, time-consuming, and hazardous. This research proposes a novel method for detecting and measuring cracks in the concrete structure of dams by utilizing a combination of the Convolutional Neural Network (CNN)-Fully Convolutional Network (FCN) algorithm and image-based 3D modeling. The method enables the identification of cracks on large structures at a pixel level, facilitating detection over a wide area and with high accuracy. Specifically, the approach detects cracks on the threedimensional surface of the dam, providing a comprehensive and effective means of identifying potential structural issues. To accurately determine the size of the detected damage, the local thickness (LT) algorithm is introduced. This algorithm utilizes the pixelprecise detection capabilities of crack detection to provide highly precise crack size measurements. The CNN-FCN system segments cracks at the pixel level on the texture space obtained from a 3D model created through photogrammetry techniques. Then the trained CNN is employed to detect crack patches, which are then imported into a trained FCN system for pixel-level segmentation. The predicted crack pixels are converted to a physical scale by the LT algorithm to quantitatively measure the morphological features of cracks. The resulting crack map is then projected onto a 3D model, providing a comprehensive visual representation of the cracks. The proposed crack detection method can achieve an accuracy of over 90%. The study revealed that the presented system could efficiently and precisely detect, locate, and measure the size of cracks in large dam structures, highlighting the potential of the proposed method to enhance the efficiency and effectiveness of crack inspection. Furthermore, the visualization of cracks on 3D models can help increase inspectors’ inspection efficiency.