DiffCNN : a neural network-based diffusion model for identification and quantification of cracks in concrete bridges

Abstract

Concrete bridges suffer degradation and damage over time, eventually leading to failure and collapse. Therefore, monitoring these structures is necessary to identify damage early and plan timely maintenance to avoid structural failures. This paper presents a neural network-based diffusion model (DiffCNN) to identify and quantify cracks in concrete bridges' deck, wall, and pavement using both transfer learning (TL) and fully trained (FT) models. The noise control and segmentation outcome are improved with a Gaussian distribution, a CNN based architecture, and a diffusion module. Experiments on the SDNET2018 dataset show that the proposed DiffCNN model achieves a crack detection accuracy of 96.85% on the bridge deck images, 94.61% on the bridge wall images, and 99.12% on the bridge pavement images in the FT mode. Furthermore, in the TL mode, the model achieved a crack detection accuracy of 99.71% on the bridge deck images, 99.62% on the bridge wall images, and 99.93% on the bridge pavement images, outperforming traditional Deep Convolutional Neural Network (DCNN) architectures and vision transformers.