This paper leverages two techniques which are transfer learning and super-resolution to reconstruct the channel state information (CSI) in the MIMO system. Knowledge transfer or transfer learning technique utilizes the pre-trained weights from different models then can be used to solve a related task which means a super-resolution technique into our models. The structure of the CSI original can be degraded once the compression process in the UE side is conducted. Therefore, by applying the knowledge transfer of the super-resolution in the reconstruction process at the BS side for boosting the detailed structure of an output CSI to be similar as an input or the original of CSI.
Additionally, the super-resolution block is split into two blocks to enhance the detailed structure of each compressed CSI. The simulation results have shown that the combination of transfer learning and split of super-resolution techniques into CSI reconstruction system (TL-SR-CSI) can achieve lower loss error which can be interpreted as higher accuracy of the CSI reconstruction against with existing system CsiNet , CSI-DO and CSI-DO-SR-1 in terms of NMSE, training, and testing loss.