Nonlocal denosing method for natural image via 2DPCA and direction information

Recently, the Nonlocal means filter have has gained wide attention. Unfortunately, the computational complexity of this methods is quite a burden. Several Speedup methods have been suggested. In our former work[17], we propose a scheme to more efficiently preselect similar patches, based on the two-dimensional principal component analysis (2DPCA). Although the method can yield good results, the computational complexity remains high. Besides, both practice and theory proved that nonlocal mean filter is suitable for processing the texture image. For natural image, its performance still needs improvement. Hence we proposed improved Semi-nonlocal version of the 2DPCA NL-mean filter, which directly employs features extracted by the 2DPCA to compute the weights, meanwhile the direction information of image patch is also used to design the weights function. Experimental results show that our method can achieve better filtering results in a variety of images, such as weak gradient image, face image and texture image.