Automatic segmentation and quantification of multiple subsurface objects in reinforced concrete ultrasonic imaging based on improved multi-task deep neural network

Ultrasonic imaging is a widely used non-destructive evaluation technique for assessing the internal condition of reinforced concrete (RC) structures. Traditional ultrasonic methods for detecting subsurface anomalies often produce complex images that are difficult to interpret, particularly by untrained personnel. This study introduces Fir-Net, a multi-task deep neural network designed to perform automatic multi-target segmentation of ultrasonic data from RC structures. Fir-Net uses a trunk-fork-branch architecture that enhances feature extraction, integrates multilevel supervision, and supports multi-task learning to automatically distinguish, quantify, and segment various subsurface objects. An RC specimen with artificial defects at various depths was prepared to create a comprehensive dataset containing ultrasonic images and ground‑truth annotations. The model processes individual B‑scans by segmenting multiple objects, and combines the segmented results into coherent three‑dimensional representations to enable accurate visualisation of the RC subsurface conditions. The experimental results demonstrate the feasibility of using Fir-Net to automatically segment multiple objects in RC.