Hybrid UNet with ResNe50 Encoder and ASPP for Accurate Road Crack Detection and Segmentation
TL;DR
A deep learning-based crack detection model using a hybrid U-Net architecture enhanced with a pre-trained ResNet50 encoder, Atrous Spatial Pyramid Pooling (ASPP), and attention gates is introduced, showing superior performance in accuracy, Dice coefficient, IoU, precision, and recall compared to traditional CNNs.
Abstract
Road surface cracks are key indicators of pavement deterioration, requiring accurate detection for timely maintenance. This study introduces a deep learning-based crack detection model using a hybrid U-Net architecture enhanced with a pre-trained ResNet50 encoder, Atrous Spatial Pyramid Pooling (ASPP), and attention gates. ResNet50 captures multi-level features, while ASPP extracts multi-scale contextual information, improving detection of cracks with diverse shapes and orientations. Attention mechanisms refine spatial features and suppress background noise, enhancing subtle crack identification. Trained on the Crack500 dataset (471 images with binary masks), the model incorporates preprocessing techniques like resizing, contrast normalization, and data augmentation to address class imbalance. Quantitative results show superior performance in accuracy, Dice coefficient, IoU, precision, and recall compared to traditional CNNs. Visual analysis confirms robustness under varied lighting and surface conditions. Future work may address challenges like noise, low contrast, and occlusion through advanced augmentation, domain adaptation, and real-time deployment via UAVs or robotic platforms.