The spectral computed tomography (CT) based on the photon-counting detector has a great potential in material discrimination for its ability to obtain the energy spectral information at multiple energy bands. Due to the poor consistency between the narrow energy-bin detection and the photon-counting detector, there are lots of noises and artifacts in the multi-spectral CT images, which is not beneficial to material decomposition and discrimination. Thus, from the point of view of reconstruction, the traditional study method based on tensor dictionary learning (TDL) is improved and a new image reconstruction method based on image total variation (TV) and TDL is developed, which is called TV+TDL for short. This method not only inherits the advantage of the TDL method in enforcing the similarity among all energy channel images, but also further recovers the slim structures and details, effectively suppresses noises, and thus improves the accuracy of material decomposition by introducing the image TV as a regularization term. The simulation results show that the TV+TDL method can effectively reconstruct high-quality multi-spectral CT images and successfully realize material decomposition and discrimination based on the base material model. The validity and practicability of this method are verified.