Automatic Multi‐label Classification of Bridge Components and Defects Based on Inspection Photographs

Abstract When performing visual inspections of bridges, experts collect photographs defects to assess the overall condition structure and schedule maintenance plans. Such are labor‐intensive, computer vision‐based systems being investigated as automated tools assist in their inspections. An important aspect however remains ensure representativeness data accounting for sheer size, complexity variety bridge components reported. Here, we perform a multi‐label classification on dataset (SOFIA dataset) that consists 139,455 images types among which 53,805 labeled (13 classes each type). The containing class imbalance noisy labeling is processed using embedding computed from unsupervised deep learning methods. A combination class‐balancing techniques state‐of‐the‐art Vision Transformer model. Interclass relations, determine whether defect should be part component, implemented with an additional filtering step. whole method also deployed CODEBRIM benchmark resulting improved accuracy score.