Automatic Concrete Crack Identification Based on Lightweight Embedded U-Net

Abstract

Crack inspection is a routine practice in maintaining and monitoring civil infrastructure. Traditional inspection methods are time-consuming, raise safety-concerning, and impose significant cost on stakeholders. Fast and accurate concrete crack detection represents a cutting-edge technology and is an essential component for automated maintenance and monitoring features. Furthermore, lightweight networks are playing a crucial role in advancing the application of automation technologies. Thus, this paper proposes a lightweight embedded U-Net (LEU-Net) that emphasizes the trade-off between high accuracy and fast inference, aiming to achieve pixel-level segmentation of concrete crack. Motivated by the superior performance of IS-Net in the dichotomous image segmentation task, LEU-Net inherits its backbone. In order to efficiently reduce computational complexity and redundancy of convolutional kernels, DWConv and SCConv replace partial normal convolution within the architecture. To enhance the extraction of crack features across diverse dimensions, a GHPA (Group multi-axis Hadamard Product Attention mechanism) module based on Hadamard Product Attention mechanism is adopted in LEU-Net. During the training process, a hybrid loss function is employed to optimize the model training and enhance generalization ability. Our experimental results demonstrate that the proposed LEU-Net can deploy in a laptop with 6GB of VRAM and achieve 62.66% and 79.38% MIoU in the bridge crack image dataset and concrete pavement crack image dataset respectively, outperforming the other state-of-the-art segmentation networks and attaining more rapid and more accuracy crack segmentation.

Keywords

• Deep learning
• crack detection
• lightweight network
• semantic segmentation
• U-Net