Less Is More? A Comparative Study of U-Net Architectures For Pavement Crack Segmentation

Automatic crack segmentation in asphalt pavement is important for road infrastructure maintenance. Convolutional neural networks, particularly U-Net-based architectures, are effective, but often trend towards increasing complexity. This study investigates if this complexity increase is justified. We compare the original U-Net, a residual connection variant (RSC-UNet), and a ConvLSTM variant (RNN-UNet) using the SUT-Crack dataset. Results show the original U-Net, after careful configuration, can outperform more complex architectures on precision and IoU. This challenges the assumption that complexity always improves results, emphasizing rigorous optimization of both architecture and hyperparameters. We also discuss the practical need for extreme pixel-level precision in various pavement repair contexts, suggesting an overly pixel-focused approach may not be optimal. The study advocates for a critical evaluation of complexity versus performance in segmentation model design, promoting parsimony.