NSCT's good multi-scale and multi-directional analysis capabilities bring effective expression of image geometry. The NSCT transformation is performed on the noisy image, the sub-band coefficients representing the image are aggregated into a geometric structure with a certain intensity, and the sub-band coefficients corresponding to the noise are randomly distributed and will not be aggregated into an effective image structure. In order to achieve image denoising, the noise coefficients should be removed as much as possible and the corresponding image coefficients should be retained as much as possible when dealing with the offspring coefficients.
The self-snake model is introduced to smooth the coefficients of the NSCT offspring, because the Self-Snake model has the ability to smooth the noise and protect the edge of the image better, which just meets the above requirements for processing the offspring coefficients.
The sub-generation coefficients are further finely divided into strong image coefficients, weak image coefficients and noise coefficients. After the entire subband is smoothed by the Self-Snake model, the amplitude of strong image coefficients is slightly weakened, the amplitude of weak image coefficients is greatly weakened, and the noise coefficient is greatly weakened but not yet completely zeroed. For this, further processing is performed for the offspring coefficient classes. Because the strong image coefficients correspond to the main content of the image, and noise pollution has little influence on them, it is considered that they should not be attenuated and need to be restored to their original values; weak image coefficients correspond to image details, and noise pollution has a great impact on them, but the Self-Snake model The protection of the geometric structure makes the noise component still remain in it, so it needs to be further attenuated; for the noise figure, its amplitude needs to be completely zeroed.
In order to realize the separate processing of different types of coefficients in the sub-bands, a threshold T is first introduced to distinguish strong image coefficients from weak image coefficients. The threshold is estimated using the BayesShrink criterion.