We address the image denosing problem in the video surveillance system in smart grid, the approach is based on compressed sensing and sparse representation theory over overcomplete dictionary. We train dictionary by noisy image, use K-SVD algorithm to update dictionary and OMP to compute sparse representation coefficients. An improved orthogonal matching pursuit algorithm based on atomic matching criterion of Dice coefficient is used to reconstruct images. Finally, we can get the denoised image. The matlab simulation experiments show that this method is an effective denoising algorithm, the denoising result for Gaussian white noise is better than the wavelet functions. Using video images which is corrupted to train the dictionary, the denoising task could be completed efficiently even there is no high-quality original image.