Based on traditional discrete total variation model, we established the separable total variation model exploiting low-dimensional projection; Combining with Frobenius norm and the convexity of image, we proposed a method that rooted in convex optimization to solve the separable discrete total variation problem, which can be applied into image denoising. Simulation results show that, with the ability of effectively keeping profile and details, the peak signal to noise ratio of 256×256 size image after denoising can reach 28.5 dB while the variance of random noise is 0.1, thus illustrating the good performance at the removal of random noise. By revising the numbers of iterations, the relationship between speed and accuracy can be balanced with considerable flexibility, thus adjusting to different denoising requirements.