A sub-pixel level algorithm in edge detection of noisy image was proposed. Firstly, the nonlinear regularized Perona-Malik model based on anisotropy diffusion was adopted to realize smooth filtering on the image. Then the improved Sobel operator was used to perform preliminary edge detection on the image and record the coordinates of the real and a few false edge points in the form of linked list with the original gray image being retained. Finally, the proposed Zernike moments were employed to locate the sub-pixel edge accurately. The results indicate that the proposed algorithm has solved the problems of excessive false edge points and the wider edge detection outcome of the traditional algorithm; the image quality treated by the proposed method is close to that treated by wavelet algorithm with a positioning accuracy of sub-pixel.