A DSF-net-based approach to dual-branch instance segmentation of weak bridge defects

Abstract

Restricted by the uniqueness of bridge structures and safety requirements of imaging devices, there are numerous challenges in computer vision-based identification of structural surface bridge damage, such as weak defects segmentation, background noise interference, and the difficulty of simultaneously detecting bridge defects with varying morphologies. To address this issue, an efficient dual-branch instance segmentation method for weak defects detection, called DSF NET, has been proposed. Considering the morphological differences among different bridge defects, the Multi-Defect Classification (MDC) Module is introduced to classify the defects into thin tubular defects, which are addressed with dynamic snake convolution(DSConv) for flexible feature extraction, and common small defects. Moreover, the problem of high proportion of weak defects and the serious imbalance of defect categories, a new multi-scale feature fusion network (IFPN) is proposed to enhance the ability to process detailed information of bridge defects. Additionally, Convolutional Block Attention Module(CBAM) was applied to enhance local features and suppress background noise and useless information. After training and optimization of DSF NET, the Mean Average Precisions for masks of tubular defects and common small defects have improved from 5.26 % to 34.92 % and 20.75–77.81 % respectively. Results demonstrate that integrating DSConv into the convolution blocks at lower stages significantly improves the performance of the model, markedly enhancing defect detection capabilities. The proposed method has the potential to provide an intelligent tool for weak bridge defect detection.