Advancing Infrastructure Integrity: Deep Learning Based Segmentation for Steel Bridge Corrosion Assessment

Corrosion presents a major risk to the structural integrity and longevity of steel bridges, necessitating effective monitoring and assessment. Unchecked corrosion can compromise structural stability and incur substantial economic and environmental costs. This research introduces a novel approach for assessing corrosion in steel bridges using deep learning-based image segmentation. Utilizing a dataset of 514 RGB images from steel bridge elements, captured by drones and Nikon cameras, and annotated at the pixel level, we evaluated various deep learning architectures. A total of 412 images were used for training, 90 for validation, and 12 for testing from the dataset. Initial experiments with U-Net, SegNet, FPN, BiSeNet, and EncNet showed promising results, with the FPN model achieving an accuracy of 86%, F1 score of 0.79, IOU of 0.65, and mAP of 0.88. Enhanced performance was achieved by incorporating VGG16 and ResNet50 backbones into the FPN model, resulting in an accuracy of 95.75%, F1 score of 0.93, IOU of 0.88, and mAP of 0.89. These findings illustrate how transfer learning enhances accuracy and reliability of deep learning-based segmentation models for corrosion assessment in steel bridge inspections.