Pavement Crack Segmentation Using an Attention-Based Deep Learning Model

It has been observed that cracks, the most common sign of deterioration happing on the pavement, are difficult to detect at an early stage. Although the U-net-based model has detected well-established cracks, it shows some limitations when working with low-quality pavement images that are automatically captured by moving pavement-inspection vehicles. In this study, the attention technique is applied to the U-net model to enhance the results of pavement crack detection under some difficult pavement image conditions. Attention gates (AGs) are deployed at the skip connections of the U-net model to remove irrelevant regions by setting attention weights for each image part. This procedure helps the U-net model learn how to eliminate extraneous regions in the input image. Therefore, the technique minimizes the computational resources by ignoring wasted irrelevant operations and enhances crack segmentation results. The proposed model is verified using a real-life image packet of pavement. The performance of the attention U-net model illustrates better outcomes compared to the ones from the U-net model.