This paper makes some improvements on Non Local Means（NL-Means) image denoising algorithm. A quantitative method is given to estimate the optimal filter parameter h. Noise variance is estimated from noise image

and h parameter is estimated from this variance and noise image standard deviation. Based on the symmetry of weighted Euclidean distance

the most complex distance computation was halved for every pair of pixels

so the computation complexity was reduced to about half of original NL-Means without performance decline. Experimental results on several images show that our adaptive non local means algorithm（ANL-Means) gives nearly best performance with only about half computation time if using original NL-Means.