Conventional practices of bridge visual inspection present several limitations, including a tedious process of analyzing images manually to identify potential damages. Vision-based techniques, particularly Deep Convolutional Neural Networks, have been widely investigated to automatically identify, localize, and quantify defects in bridge images. However, massive datasets with different annotation levels are required to train these deep models. This paper presents a dataset of more than 6900 images featuring three common defects of concrete bridges (i.e., cracks, efflorescence, and spalling). To overcome the challenge of limited training samples, three Transfer Learning approaches in fine-tuning the state-of-the-art Visual Geometry Group network were studied and compared to classify the three defects. The best-proposed approach achieved a high testing accuracy (97.13%), combined with high F1-scores of 97.38%, 95.01%, and 97.35% for cracks, efflorescence, and spalling, respectively. Furthermore, the effectiveness of interpretable networks was explored in the context of weakly supervised semantic segmentation using image-level annotations. Two gradient-based backpropagation interpretation techniques were used to generate pixel-level heatmaps and localize defects in test images. Qualitative results showcase the potential use of interpretation maps to provide relevant information on defect localization in a weak supervision framework.

Bridges deteriorate over time, which requires the continuous monitoring of their condition. There are many digital technologies for inspecting and monitoring bridges in real-time. In this context, computer vision has extensively studied cracks to automate their identification in concrete surfaces, overcoming the conventional manual methods that rely on human judgment. The general framework of vision-based techniques consists of feature extraction using different filters and descriptors and classifier training to perform the classification task. However, training can be time-consuming and computationally expensive, depending on the dimension of the features. To address this limitation, dimensionality reduction techniques are applied to extracted features, and a new feature subspace is generated. This work used histograms of oriented gradients (HOGs) and uniform local binary patterns (ULBPs) to extract features from a dataset containing over 3000 uncracked and cracked images covering different patterns of cracks and concrete surface representations. Nonlinear dimensionality reduction was performed using kernel principal component analysis (KPCA), and three machine learning classifiers were implemented to conduct the classification. The experimental results show that the classification scheme based on the support-vector machine (SVM) model and feature-level fusion of the HOG and ULBP features after KPCA application provided the best results as an accuracy of 99.26% was achieved by the proposed classification framework.