Hyperspectral images contain rich spectral information, and the hyperspectral image reconstruction from a single RGB image is of great value to military target recognition and medical diagnosis. Since traditional algorithms cannot reconstruct RGB images with unknown spectral response from cameras, this paper proposes a reconstruction algorithm based on an improved residual dense network. First, with an improved residual dense block as the basic module, we apply the adaptive weight module for feature recalibration, which improves the accuracy of hyperspectral image reconstruction. Additionally, our algorithm solves the hyperspectral image reconstruction instead of image super-resolution through replacing the spatial transformation layer with a feature one, which transforms the network from the spatial dimension to the spectral dimension. The experimental results show that the proposed algorithm is superior to the traditional methods and deep learning methods in both subjective effect and objective evaluation indicators. Compared with those of the sparse dictionary method, the mean relative absolute error (MRAE) and root mean square error (RMSE) of the proposed algorithm are reduced by 46.7% and 44.8%, respectively.