For NeighShrink method used in the image de-noising, a new image de-noising algorithm is proposed to keep image edges more effectively, and it mainly improve the domain of NeighShrink which is fixed. The new method can segment the image into many domains adaptively to de-noise the images. Furthermore, combined with wavelet correlation in the same layer, we get various irregular neighborhoods with a fixed window, and choose the coefficients which have closer geometric distance and are in the same irregular neighborhood to improve NeighShrink. This method decomposes noisy images with stationary wavelet transform to keep phase invariance. Then, in accordance with special rules, it segments the low frequency sub-band by using Pulse Coupled Neural Networks (PCNN) model, and then gets the approximate information. And the edge information will be protected during the de-noising process. A better restoration of images is demonstrated in the results of experiments, with detail of images kept as well as image noises decreasing.