A multilevel bridge corrosion detection method by transformer-based segmentation in a stitched view

Abstract Corrosion is one of the main damage mechanisms in civil engineering structures today. Rapid identification and accurate assessment of corrosion in structures are essential to ensure the efficient allocation of limited funds for the maintenance and renewal of existing structures. Vision-based neural networks have been widely used in corrosion detection, in which convolutional neural network (CNN)-like models remain dominant. However, these conventional network models exhibit a saturating performance. Because of the self-attention mechanism, the transformer is the newest breakthrough in computer vision and is becoming state of the art. As the complexity of structures increases, transformer-based methods have no saturating performance. This study proposes a corrosion localization and evaluation architecture for a larger view based on semantic segmentation and image stitching for automatic localization and diagnosis of corrosion from stitched images. The experimental results showed that the proposed method achieved better corrosion detection performance (F1-score = 68.2%) than that of the mainstream CNN-like models U-Net (F1-score = 61.8%) and DeepLabV3 + (F1-score = 60.1%). Image stitching is utilized for corrosion segmentation in larger view images, and the field test shows that the proposed architecture could stitch corrosion prediction from different images.