Inspecting concrete bridge decks is essential for ensuring infrastructure safety and durability, yet traditional manual inspection methods are labor-intensive, time-consuming, and prone to human error, creating demand for automated solutions. This research develops an automated bridge deck inspection framework that combines Unmanned Aerial Systems (UASs) with advanced Deep Learning (DL) techniques. The framework uses UAS to capture high-resolution optical images of a concrete bridge deck, which are then analyzed using a novel Residual Vision Transformer (RvT) architecture featuring convolutional token embedding and depthwise residual projection. The RvT model was systematically compared against four established architectures: Vision Transformer, Convolutional Vision Transformer, ResNet50, and traditional Convolutional Neural Network using stratified train–validation–test splits and rigorous hyperparameter optimization. Following promising preliminary results, the proposed RvT model was subjected to a more stringent fivefold stratified cross-validation to rigorously assess its stability and generalization capabilities. This robust validation confirmed its consistent high performance and ability to avoid overfitting. Experimental evaluation demonstrated that the cross-validated RvT model achieved a robust and reliable performance, attaining a final test accuracy of 95%. While its overall accuracy was competitive with other leading models, the RvT distinguished itself by achieving the highest precision (0.92) on the minority crack class. These findings highlight the effectiveness of hybrid transformer-based architectures that combine convolutional inductive biases with self-attention mechanisms for infrastructure defect detection. By integrating advanced DL models with UAS-collected imagery, this approach offers significant potential to reduce inspection times and costs while improving accuracy and reliability for transportation agencies and DOTs, supporting more efficient bridge maintenance and safety management.
Khatry et al. (Fri,) studied this question.