Adptive Decomposition For Edges And Textures Based On Modeling And Directional Filtering

It is well known that human are sensitive to regions where grayscale changes rapidly, so regions full of textures and edges are of importance in image processing. From the perspective of energy, however, most energy centralizes in regions with low frequency. Previous literatures can be categorized into four classes which are signal processing, computer graphics, fractal, and random fields. Among all of these methods, signal processing and random fields have been attached the most attention. Ridgelet, Curvelet, and Contourlet, belonging to signal processing methods, choose different basis functions for image decomposition.Being quite different from previous signal processing methods, this paper proposed a brand new dual multiresolution decomposition through introducing orientation multiresolution and orientational filter bank based on texture. This method can not only capture linear singularities which wavelet is not good at, but it has better compression performance comparing with traditional methods. Firstly, wavelet is utilized to realize spatial multiresolution, which can avoid redundancy brought by LP decomposition. Secondly, orientation distribution is predicted respectively in three high frequency subbands of wavelet. Invoked by motion compensation method in all-phase wavelet, orientation prediction is based on all-phase wavelet decomposition in order to avoiding texture damaging brought by phase loss. Thirdly, orientational filter bank is designed according to the prediction of orientation distribution, and these subbands are non-uniform which embodies orientational multiresolution in the framework. In order to simplify the design of non-uniform orientational filter bank, it is approximated by changing the topology of non-symmetric binary tree filter bank.The adaptive dual multiresolution decomposition designs appropriate basis functions for different images to meet different orientation distributions of textures. These basis functions are not only anisotropic, but their spatial distributions are also non-uniform. This non-uniform orientational basis function is much more relevant to the original image, leading to better energy convergency and sparser coefficients. The testing result of this new framework manifests much better performance on PSNR and MSE in comparison to the traditional.