PSH-UNet: a parallel state-hybrid lightweight network for efficient crack segmentation

View at publisher Summarise with VIRAT
Authors: Yilei Shu, Caixia Wang, Peifang Xin, Li Wang, Shuai Jiang, Longwang Yue
Cited by: 0
References: 26

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Abstract Accurate crack detection and segmentation are crucial for monitoring the structural safety of buildings; however, existing deep learning methods frequently encounter challenges, including low computational efficiency and incomplete crack recognition. To address these limitations, a lightweight crack segmentation network, denoted as PSH-UNet, is proposed. The core of the proposed architecture lies in the designed Parallel-State-Hybrid (PSH) block, which adopts a gated parallel architecture. This architecture integrates the Local-PSH (L_PSH) module, utilized for the extraction of subtle local features, with an S_PVM layer based on State Space Models. This integration enables the network to maintain efficient computational performance while preserving long-range dependencies within images. Extensive experimental evaluations were conducted on two public datasets: Crack500 and DeepCrack. The results demonstrate that PSH-UNet achieves superior performance compared to advanced baseline models, reaching an mIoU of 83.7% on the DeepCrack dataset. Notably, this accuracy is achieved with a parameter count of merely 4M, reducing model complexity by over 90% when compared to high-performance baseline models. These findings demonstrate that the design of PSH-UNet achieves an optimal balance between segmentation accuracy and computational efficiency, providing a novel solution for the deployment of crack detection systems on edge devices constrained by limited computational resources.

Fields of Study & Subjects

Infrastructure Maintenance and Monitoring Advanced Neural Network Applications Industrial Vision Systems and Defect Detection

Publication details

Journal / source
Engineering Research Express
Volume
8
Issue
7
Pages
075229–075229
Publication date
26 March 2026
Language
EN