Deep learning-based segmentation model for permeable concrete meso-structures

Abstract

The meso-structure of pervious concrete significantly influences its overall performance. Accurately identifying the meso-structure of pervious concrete is imperative for optimizing the design of pervious concrete, considering its mechanical properties and functionality. Therefore, to address the difficulty of recognizing the meso-structures of pervious concrete, a method utilizing deep learning image semantic segmentation techniques is proposed in this study. First, based on the classical deep learning model, three models, namely, Res-UNet, ED-SegNet, and G-ENet, are proposed for recognizing pervious concrete meso-structure using deep learning image semantic segmentation techniques. These models introduce a residual module, a hybrid loss function, and a differential recognition branching structure to enhance the ability to recognize detailed information within pervious concrete meso-structure and small targets. Second, the respective recognition performances of these methods on the meso-structure of pervious concrete were thoroughly analyzed by experiment. The results indicate that the proposed three recognition methods for recognizing the meso-structure of permeable concrete outperform conventional techniques not only in terms of efficiency but also in recognition accuracy and the ability to distinguish and identify aggregates, pores, and cement binders. In terms of comprehensive recognition effectiveness, the Res-UNet model outperforms, followed by ED-SegNet and G-ENet. Furthermore, the computational efficiency of these three recognition methods meets the requirements of engineering applications.