Abstract: An improved non-local means（NLM） image denoising algorithm is proposed, which uses Mahalanobis distance to measure the similarity between the image pixels. Firstly, calculating the Mahalanobis distance between the image pixels in the eigenspace since the Mahalanobis distance is not robust in the sample space. Secondly, the image data is analyzed with the principal component analysis method, thus the Mahalanobis distance equation is simplified. Finally, the improved NLM image denoising algorithm is obtained with the Gaussian weighted kernel function which is composed of the simplified Mahalanobis distance. The experimental results on several typical images show that the improved NLM algorithm can achieve better denoising effect than the original NLM algorithm with a variety of image quality evaluation method. The filter parameter ‘h’ in the improved NLM denoising algorithm is analyzed in details and the equation between the filter parameter ‘h’ and the image noise variance is estimated. Based on the equation, the experimental results achieve nearly best denoising performance of the improved filtering algorithm.