Intelligent Identification of Rivet Corrosion on Steel Truss Bridge by Single-Stage Detection Network

Abstract

Rivet corrosion, which is a common disease of steel truss bridges, directly reflects the safety status of steel structures. The identification of rivet corrosion is critical to ensure the normal service of steel truss bridges. In practical engineering, the main detection method of rivet corrosion is manual visual inspection. However, this method has low efficiency and poses a threat to the personal safety. To address this issue, an intelligent identification method for rivet corrosion on steel truss bridges by a single shot detector (SSD) is proposed after obtaining the panoramic image of the bridge. The sub-images cut from the panoramic image are as the networks input. Considering the small size of bridge rivets and low precision of small object detection of SSD, this study divides the panoramic image into sub-images of 100 — 100 pixels, and then uses bilinear interpolation to resize the sub-images into 300 — 300 pixels. To improve the robustness of the detection model, gaussian noise, random rotation and roll-over tricks are applied to the original dataset. The expanded dataset includes 600 labelling images, which is divided into training set (80%) and testing set (20%), including corroded rivets and normal rivets. The network is trained with transfer learning technique for 12000 iterations, with cross entropy loss for classification and smooth L1 loss for location. The confidence threshold in network inference is set to 0.6 considering the rivet space distribution to reduce false positives of corroded rivets. The qualitative and quantitative testing results show the accuracy of the proposed approach.