Multispectral reconstruction from single RGB image can eliminate the geometric distortion problem existing in optical bandpass filters-based multispectral cameras and save the capturing time, but the reconstruction accuracy is limited by just using the original three channels response. Camera response expansion is an optimal choice to increase the dimensions of the input information for multispectral reconstruction as the imaging and processing is practically not linearity for the trichromatic digital cameras. In this paper, the camera response expansion based on polynomial model was tested for multispectral reconstruction from single RGB image, the pseudoinverse method was adopted for the training-based multispectral reconstruction, and the local inverse distance weighted (LIDW) optimization was proposed to improve the reconstruction accuracy. The proposed method was compared with the current existing methods through practical experiment, and the results indicated that it outperformed existing methods.