A maximum likelihood estimation image denoising method is proposed using the non-local principle component analysis. Pixels with high similarity in both the gray level and the texture information are selected, and used to implement the maximum likelihood estimation. This kind of optimal restored method can overcome the drawback of the local image denoising method such as blurring edge, and improve the accuracy for restoring detail information in image using maximum likelihood estimation. Experiments using the proposed method, principal neighborhood dictionary non-local mean method and local maximum likelihood estimation method are implemented in images with different noise standard and different geometric complexity, and the performance of aboving three denoising methods are compared quantitatively and qualitatively. The results demonstrate that the proposed method can remove the noise effectively and preserve detail imformation of images compared with the currently used methods.