关键词: 水工混凝土, 裂缝检测, 语义分割, 特征量化, 深度学习

Abstract: Concrete cracking is common in the main structures of hydraulic buildings, and crack detection is always crucial to hydraulic engineering safety appraisal. Shape segmentation and feature quantification are the main tasks of digital image processing technology that has been widely used in structural surface crack detection due to its advantages of high efficiency and low cost, but traditional image processing has the shortcomings of more manual intervention and weaker generalization ability. This paper describes a new method for pixel-level shape segmentation and feature quantification of hydraulic concrete cracks based on deep convolutional neural networks. This method is on the basis of U-net semantic segmentation architecture and incorporates the transfer learning technology. Specifically, it uses a pre-trained VGG16 network to enhance the encoder and extract multi-scale and high-level semantic information, and alleviates class imbalance problems by improving the cross-entropy loss function, so that the crack shape can be accurately segmented. Then, we present a set of algorithms based on the binarized segmentation mask and computer vision technology for calculating key geometric parameters such as crack area, length and width. To verify and evaluate this crack detection method, we generate an image dataset of hydraulic concrete cracking through numerical simulations and conduct comparative experiments demonstrating its effectiveness and superiority. The results indicate that its segmentation effect is significantly better than that of classical image segmentation methods, and its calculations of crack feature parameters meet the required detection accuracy. Thus, it is a new useful technique for quality control of hydraulic concrete structures.

Key words: hydraulic concrete, crack detection, semantic segmentation, feature quantification, deep learning