Due to the discarding of downlink channel state information (CSI) amplitude and the employing of iteration reconstruction algorithms, 1-bit compressed sensing (CS)-based superimposed CSI feedback is challenged by low recovery accuracy and large processing delay. To overcome these drawbacks, this letter proposes a fusion learning scheme by exploiting the bi-directional channel reciprocity. Specifically, a simplified version of the conventional downlink CSI reconstruction is utilized to extract the initial feature of downlink CSI, and a single hidden layer-based amplitude-learning network (AMPL-NET) is designed to learn the auxiliary feature of the downlink CSI amplitude. Then, based on the extracted and learned amplitude features, a simple but effective amplitude-fusion network (AMPF-NET) is developed to perform the amplitude fusion of downlink CSI and thus improves the reconstruction accuracy for 1-bit CS-based superimposed CSI feedback while reducing the processing delay. Simulation results show the effectiveness of the proposed feedback scheme and the robustness against parameter variations.