Surface damage detection for concrete bridges using single-stage convolutional neural networks

Detecting surface damages is vital for maintaining the health and safety of concrete bridges. Currently, most of image-based detection techniques are based on handcrafted low-level features which make them less applicable to actual images taken under varying environmental conditions. Recent rapid advancement in convolution neural network has enabled the development of deep learning-based visual inspection techniques for detecting multiple structural damages without needing manually-crafted features. However, most deep learning-based techniques are built on two-stage, proposal-driven detectors using less complex image data, which is not promising for practical applications and for integration within intelligent autonomous inspection systems. In this study, a faster, simpler single-stage convolutional neural network is proposed based on the real time object detection technique, You Only Look Once (YOLOv3) for detecting multiple surface damages on concrete bridges. A large field inspection images dataset of bridge damage is used for training and testing of YOLOv3. The results show that the trained YOLOv3 has a detection accuracy of up to 66%.