Improved structure-adaptive anisotropic filter based on a nonlinear structure tensor

A variety of structure-adaptive filters are proposed to overcome the blurred effects of image structures caused by the classical Gaussian weighted mean filter. However, two major issues are needed to be dealt with carefully for structure-adaptive anisotropic filters. One is to properly construct the filter kernel and the other is to accurately estimate the orientation of the image structures. In this paper we propose to improve the structure-adaptive anisotropic filtering approach based on the nonlinear structure tensor (NLST) analysis technique. According to the anisotropism measurements of image structures, a new kernel construction method is designed to make the filter shape fine adapted to image features. Through the accurately estimated orientation of the image structures, the filter kernels are then properly aligned to perform the filtering process. Experimental results show that the proposed filter denoises the noisy images carefully and image features, such as corners and junctions are well preserved. Compared with some other known filters, the proposed filter obtains great improvements both in Mean Square Error (MSE) and visual quality.