Interscale image denoising with wavelet context modeling

The paper presents a wavelet-based linear minimum mean square-error estimation (LMMSE) scheme to exploit the strong wavelet interscale dependencies for image denoising. Using overcomplete wavelet expansion (OWE), we group the wavelet coefficients with the same spatial orientation at adjacent scales as a vector. The LMMSE algorithm is then applied to the vector variable. This scheme exploits the correlation information of wavelet scales to improve noise removal. To calculate the statistics of wavelet coefficients more adaptively, we classify them into different clusters by the context modeling technique, which yields a good local discrimination between edge structures and backgrounds. Experiments show that the proposed scheme outperforms some existing denoising methods. A biorthogonal wavelet, which well characterizes the interscale dependencies, is found very suitable for the scheme.