Gradient Transformer Self-Attention U-Net: Enhancing Crack Detection in Concrete Bridges for Sustainable Maintenance

Abstract Accurate identification of cracks in concrete bridge components is vital for infrastructural maintenance and ensuring public safety. This paper presents the gradient transformer attention U-Net (GTAU-Net) model, which processes input images via quantum-inspired hybrid filtering for effective classification of cracks in bridge components. Based on quantum physics principles, highly probable crack areas in the form of multichannel features, such as edges and textures, are amplified to generate quantum fused filter (QFF) images. GTAU-Net processes these QFF images to compute gradient saliency scores (GSSs) before significance embedding within the feature map. The GSSs contribute to the adaptive attention mechanism, which allows the model to not only process all input tokens but also selectively focus on the most relevant ones, thereby improving both computational efficiency and interpretability. Based on experimental results, GTAU-Net achieved 99.42% accuracy in classifying cracks in bridge components, outperforming existing models. Additionally, it measured the crack lengths accurately, which highlights its potential utility in sustainable maintenance.