The conventional wavelet transform technique is efficient when dealing with 1-D signals. But for 2-D images with abundant geometrical texture information, the standard 2-D separable wavelet transform is not satisfying because of its directional absence. In order to obtain better image representation, we present a novel image de-noising method in terms of the intrinsic geometrical characteristics of images. Based on the principle of the minimal approximation error, we search for geometrical directions in each initialized direction windows, and combine smaller windows with a bigger one according to the regulation of combination. When the best denoising direction is obtained, we project 2-D information on the 1-D along the certain directions and perform 1-D wavelet transform for denoising. The validity and efficiency are shown by experiments.
