The work aims to effectively filter noise in steel strip image and retain more image details, provide a good source figure for strip image-based surface defect detection. The strip image was decomposed to high and low frequency coefficients of the image as Nonsubsampled Contourlet Transformation (NSCT) had characteristics of fine decomposition. On the basis of maximum posterior probability, a self-adaptive threshold model was established for pre-denoising with variance of noise image coefficient. In order to further remove noise in steel strip image, the pre-denoised steel strip image was processed in the method of non-local means, and noise in the steel strip image was filtered effectively in this way. The simulation results showed that, compared with control group method, the proposed method was free from staircase effect and other shortcomings, and generated higher peak signal-to-noise ratio and structural similarity. The proposed method can effectively remove noise in the steel strip image, and can retain more image details, so that the denoised image has better visual effect.