A novel deep learning method for automated microcrack segmentation in cement-based materials

Abstract

The presence of microcracks significantly degrades the mechanical and durability performance of cement-based materials. Scanning electron microscope (SEM) images can clearly reveal microcracks, but the current mainstream microcrack segmentation methods are semi-automatic and limited to specific databases or conditions. To address these challenges, this study first constructed a microcrack dataset containing SEM images with various microcrack characteristics. Subsequently, a high-quality feature extraction and fusion network for microcrack segmentation (QF-Net) was proposed. The network incorporates two customized components aimed at improving segmentation accuracy, with CNN-Trans-Fusion block for feature extraction and Haar wavelet and max pooling hybrid downsampler (HWMD) for downsampling. Through training on the microcrack dataset with a combined loss function, QF-Net forms its corresponding network model. Experimental results demonstrate that the QF-Net model achieves traditional methods and state-of-the-art networks, and also shows remarkable generalization ability on images from published studies. Ablation studies confirm the significant contributions of the introduced modules to the overall segmentation performance. This study provides an automated and accurate method for microcrack segmentation, giving insights for understanding the formation mechanisms of microcracks and optimizing material properties during design.