BiNet: Bridge Visual Inspection Dataset and Approach for Damage Detection

Manual damage identification from large visual inspection data sources demands tremendous effort and is prone to discrepancies due to human errors, fatigue, and poor judgments of bridge inspectors. Deep learning techniques have obtained state-of-the-art results in solving computer vision tasks across different domains such as health, retail, among others. To encourage the development of automated visual inspection and damage detection solutions in the realm of infrastructure management, we propose BiNet, a visual inspection dataset for multi-label damage identification that can be used for classification, localisation, and object detection. We have investigated and compared the performance of convolutional neural networks and transfer learning approaches for automated damage classification and localisation. We have established baseline performance results of BiNet for future comparisons. Our contribution is introducing the public well-curated bridge visual inspection dataset and a deep learning approach for automated damage detection. This work is a step toward (semi) automated inspection of bridge structures for cost-effective, consistent and reliable bridge management.