# DSCformer: A Dual-Branch Network Integrating Enhanced Dynamic Snake Convolution and SegFormer for Crack Segmentation

**Authors:** Kaiwei Yu, I-Ming Chen, Jing Wu  
**arXiv:** [2411.09371](https://arxiv.org/abs/2411.09371)  
**Date:** 2024-11-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2411.09371v1/

## Key points

- Accurate detection and segmentation of cracks in concrete structures are crucial for safety and maintenance
- Current CNNs show impressive performance but struggle with intricate backgrounds and detailed tubular structures in cracks
- DSCformer is a hybrid model combining enhanced DSConv with Transformer architecture for crack segmentation tasks
- Enhanced DSConv component includes pyramid kernel for adaptive offset computation and bi-directional learnable offset iteration
- Introduction of Weighted Convolutional Attention Module (WCAM) refines channel attention for more precise feature attention
- DSCformer achieves high IoU scores of 59.22% and 87.24% on Crack3238 and FIND datasets, outperforming state-of-the-art methods
- Promises to enhance safety measures and maintenance practices in construction quality monitoring

## Summary

In the realm of construction quality monitoring, the accurate detection and segmentation of cracks in concrete structures are crucial for ensuring safety and proper maintenance. Current convolutional neural networks (CNNs) have shown impressive performance in crack segmentation tasks, but they often struggle with intricate backgrounds and fail to fully capture the detailed tubular structures present in cracks. To address these challenges, a team of researchers led by Kaiwei Yu, I-Ming Chen, and Jing Wu introduces DSCformer: a groundbreaking hybrid model that combines an enhanced Dynamic Snake Convolution (DSConv) with a Transformer architecture specifically designed for crack segmentation tasks. The key innovation lies in the enhanced DSConv component, which incorporates a pyramid kernel for adaptive offset computation and enables simultaneous bi-directional learnable offset iteration. These enhancements significantly boost the model's performance in capturing intricate crack patterns with precision. Additionally, the researchers propose a Weighted Convolutional Attention Module (WCAM) that refines channel attention within the model, allowing for more precise and adaptive feature attention. Through rigorous evaluation on two benchmark datasets - Crack3238 and FIND - DSCformer achieves impressive Intersection over Union (IoU) scores of 59.22% and 87.24%, respectively. The experimental results clearly demonstrate that DSCformer outperforms state-of-the-art methods across various datasets, showcasing its efficacy in crack segmentation tasks within concrete structures. This innovative approach holds great promise for enhancing safety measures and maintenance practices in construction quality monitoring processes.