Ground Penetrating Radar (GPR) is one of the most important non-destructive evaluation (NDE) devices to detect the subsurface objects (i.e. rebars, utility pipes) and reveal the underground scene. Due to the background noise and unknown dielectric property of GPR data, the reconstruction of the subsurface objects with metric information remains the main challenge in GPR inspection. In this letter, a novel metric GPR migration method is proposed consisting of three steps to address the above challenge. Firstly, a noise removal segmentation model is introduced to clear the GPR raw data; furthermore, a DielectricNet is proposed to predict the dielectric of each GPR B-scan data and the depth of a target object; at the last step, we optimize the migration algorithm by using random motion pattern, thus to obtain the 3D metric migration result. We use both the field and synthetic data to verify the proposed method, experimental results show that our proposed method is able to achieve at least 10% improvement in reconstruction accuracy compared with the conventional migration methods.