Road Crack Segmentation Based on Dual Branch Attention Mechanism and Multi-Scale Features

In road surface crack segmentation algorithms, obtaining comprehensive contextual information is crucial. While many solutions use the Transformer architecture for global information, its computational complexity requires combining it with CNN for feature extraction, making multi-scale contextual information extraction challenging. Although UNet-based networks are popular for fine-grained features, their decoder introduces redundant contents. To address above-mentioned issues, The paper proposes DBAUNet, incorporating SVD-ASPP (Same Value Dense ASPP) for global information extraction and DBAM (Dual Branch Attention Mechanism) to enhance crack information while suppressing background information. SVD-ASPP prevents crack disconnections, ensuring more accurate segmentation results.