In traditional Non-Local Means (NLM) algorithm, the weighted Euclidean norm can not truly reflect the similarity between two neighborhoods under large noise standard deviation. To address this problem, a new NLM denoising algorithm combined with similarity weight was proposed. Firstly, the noise image was decomposed by using the advantages of stationary wavelet transform, and the filtering function was used to predenoise each detailed subband data. Secondly, according to the refined image, the similarity reference factor between the patches was calculated, and it was used to replace Gauss kernel function of the traditional NLM algorithm. Finally, to make the similarity weights more in line with the characteristics of Human Visual System (HVS), the block Singular Value Decomposition (SVD) method based on image structure perception was used to define neighborhood similarity measure, which can more accurately reflect the similarity between neighborhoods compared with the traditional NLM. The experimental results demonstrate that the hybrid similarity weighted NLM algorithm performs better than the traditional NLM in retaining the texture details and edge information, and the Structural SIMilarity (SSIM) index measurement values is also improved in comparison with the traditional NLM algorithm. When the noise standard deviation is large enough, the proposed approach is of effectiveness and robustness.