Aiming at the problem of over enhancement and low detailed in traditional image enhancement algorithm, an infrared image enhancement method based on Retinex theory and probability nonlocal mean is proposed. Firstly, the grayscale in deep dark and bright parts of image is adjusted by the single scale Retinex method. Then the image is decomposed into basic level and detail level by probability nonlocal mean filtering. For the basic layer, histogram equalization is used to stretch contrast, and the nonlinear function is used to enhance details for the detail layer. Finally, the different levels of the enhancement results are fused to obtain the infrared image with the contrast and detail enhanced. The simulated experiments on infrared images of different scenes are carried out through the proposed method. And the results are compared with those of different enhancement algorithms on the subjective and objective sides. The comparisons demonstrate that proposed method has better results in detail and contrast enhancement of infrared image.