Internal structure recognition of EPS composite soil using fully convolutional network

Abstract

As a geotechnical backfill material, expanded polystyrene (EPS) composite soil has the advantages of low weight, high strength and easy in-site handling. EPS composite soil exhibits heterogeneity in mechanical behavior mainly due to the non-uniform distribution of EPS beads in the soil. In order to recognize the internal structural characteristics of EPS composite soil, this study applies the fully convolutional network (FCN) to identify each individual EPS bead on the cross-section planes of EPS composite soil samples. FCN is a powerful deep learning architecture that can make pixel-level classification of images. This paper introduces the structure of the FCN model, the data preparation and model training for the EPS bead identification task. The trained FCN model can satisfactorily identify EPS bead locations and in most cases give precise boundaries of EPS beads. However, for two adjacent EPS beads, the model may predict artificial overlap of the two EPS beads. Overall, the FCN model is capable of recognizing and assisting quantitative assessment of the internal structure of EPS composite soil.