Image reconstruction with large scale of fluorescence molecular tomography (FMT) data needs a large amount of computational memory and time. In order to reduce the ill-posedness of FMT and speed up the reconstruction，an accelerated reconstruction method for FMT based on locality preserving projections (LPP) and sparse regularization algorithm is presented, combining with the manifold learning and compressive sensing theory. The original fluorescent multi-projection data is reconstructed. Simulation experiment of non-homogeneous cylinder single and double goal and real experiment are performed to evaluate the reconstruction effect and time of the proposed method respectively. Experimental results demonstrate that the proposed method can ensure the accuracy and resolution of the FMT reconstruction image and reduce the reconstruction time greatly.