In order to sufficiently exploit the image information residing in the residual image for boosting the denoising performance of the Non-local Means (NLM) algorithm, a novel multi-stage residual filtering method is proposed. Firstly, the Non-Local Means algorithm is applied to a noisy image to produce an initial denoised image and a weight distributing matrix. Then the fixed-weight NLM algorithm is applied to the residual image followed by a Gaussian filtering process, which can extract the image content out from the residual as a compensation image. The compensation image is then added back to the denoised image to generate an enhanced restored image. An iterative scheme, whose principle and feasibility are derived and proved theoretically, is developed for the above filtering procedure; meanwhile a novel stopping criterion with no reference image required is proposed to determine the optimal number of iterations adaptively. Experimental results demonstrate that the proposed stopping criterion behaves similarly as the PSNR rule, and compared with the original NLM approach, the proposed method can boost the denoising performance significantly with 1.2 dB PSNR gains achieved on average and more detail information preserved, while the computational complexity is not apparently increased.