Computational study demonstrates improved road crack segmentation using a hybrid U-Net with ResNet50 and ASPP, indicating enhanced automated pavement maintenance.
Key Points
To develop an accurate deep learning framework combining a U-Net architecture, ResNet50 encoder, ASPP, and attention gates for automated road crack detection and segmentation.
Constructed a hybrid U-Net model integrating a pre-trained ResNet50 encoder for feature extraction, Atrous Spatial Pyramid Pooling (ASPP) for multi-scale context, and attention gates for spatial refinement.
Trained and evaluated the network on the Crack500 dataset (N=471 images with binary masks) using image resizing, contrast normalization, and data augmentation.
Achieved higher accuracy, Dice coefficient, IoU, precision, and recall than traditional convolutional neural network models.
Demonstrated qualitative robustness in segmenting subtle and irregularly oriented cracks across diverse lighting and pavement surface conditions.