Abstract: The purpose of ground penetrating radar signal inversion is to improve the accuracy of data interpretation. In the process of GPR signal inversion interpretation, the traditional inversion method faces the problems of time-consuming inversion process and the inversion results are greatly affected by the subjective experience of professionals. A data-driven deep learning inversion network, GInet, is proposed in this paper, which integrates multi-scale convolution and spatial attention mechanisms to invert GPR signals. Firstly, the B-Scan data dimension transformation unit is used to compress and extract features of GPR image data, and then the multi-scale UNet unit is used to convert the GPR image to the corresponding underground dielectric constant distribution image. The simulation data set built based on the finite-difference time domain method is used for network training and testing. The mean square error of GInet inversion results is reduced to 0. 005 54, and the structural dissimilarity value is 0. 000 026 7. The measured data of sand tank are used to test GInet, and the relative permittivity image obtained by inversion is close to the real model, indicating that this method can accurately and efficiently complete the GPR image inversion task.