Abstract: Considering extant method inaccurately selects the foreground region as the template of sparse reconstruction algorithm leading to a deviation for detecting a salient object, this paper presents a method of optimized background template. The first step is to calculate the boundaries of the connectivity of multiple regions and the image's templates are decided whether to belong to the background; then each region of the whole image is reconstructed by a dictionary with all the background templates. This process uses a novel sparse weighted method suppression of non zero vector matrix, so as to enhance the solution vector, which plays an important role between the templates. Finally, the final saliency map is generated by the calculation of the reconstruction error of each region. Experiments on the three standard datasets show that our optimization algorithm can effectively improve the accuracy of the algorithm and also generate a clear saliency map even in a complicated background. Compared with the original algorithm, the mean absolute error is reduced by nearly 20%.