Abstract: Aiming at the limitations of the existing ground penetrating radar inversion methods based on neural network, such as incomplete feature extraction, low accuracy and poor spatial correspondence, based on the encoder-decoder structure a new ground penetrating radar inversion network—attention inversion network (AINet) is proposed, which integrates multi-head attention mechanism and residual structure. Firstly, single dimensional convolution and residual structure are used to extract and compress effective features; then multi-head attention mechanism is used to extract global semantic information of B-scan, which improves the spatial correspondences between data pairs and optimizes relevant features; finally, the reconstruction of the dielectric constant graph is completed through a dielectric constant decoder. In the experiment, the AINet is trained, verified and tested by the ground penetrating radar inversion data set based on the finite difference time domain method, and the dielectric constant distribution map is successfully reconstructed. Compared with the existing deep learning inversion methods, the mean square error, mean absolute value error and structural dissimilarity index of AINet are significantly lower than those of other networks in a variety of generalization scenarios. The inversion test of AINet is carried out by using measured data, and the results show that the proposed method can efficiently and accurately inverse the distribution of underground dielectric constant according to the measured data of ground penetrating radar.