An Image Enhancement Algorithm Based On Contourlet Transform

Image enhancement is a technique which can improve visual effect by enhancing some information and suppressing others. The spatial domain method and the transform domain method are two traditional image enhancement algorithms. In spatial domain, histogram equalization and sharpening processing are two methods used widely, while in transform domain, there are many algorithms, including filtering and algorithms based on wavelet, curvelet, contourlet etc.Contourlet transform aims at obtaining sparse presentations of images, which inherits local supporting and the multi-resolution property of wavelet. Furthermore, it is anisotropic and provides more directions than wavelet. However, contourlet transform is not shift-invariant and could cause pseudo-Gibbs phenomena around singularities in image enhancement. Nonsubsampled contourlet transform (NSCT), as a fully shift-invariant form of contourlet transform, leads to better frequency selectivity and regularity. Thus, NSCT is unitized as the multiscale transform in image enhancement in this paper.In transform-based image enhancement algorithms, a threshold is firstly selected to distinguish noise and edge, followed by an enhancement function to amplify the useful information, e.g. edges and textures. Thus, in this thesis, distribution of NSCT coefficients in different directions and scales is analyzed, which is non-gaussian, high-peak and long-trailing, and modeled by Generalized Gaussian Distribution (GGD). In addition, an adaptive bayes threshold is constructed to suppress the noise and preserve the texture. In order to estimate the noise variance, white gaussian noise is decomposed by NSCT and the variance in different scales is found to obey negative exponent rule. Then, a fast and simple enhanced function is proposed to overcome the shortcoming of piecewise nonlinear operator, also parameters of our method which could affect the enhancement results are discussed. Finally, DV-BV objective evaluation is introduced to depict the detailed variance and background variance respectively.Experimental results demonstrate that the proposed algorithm outperforms histogram equalization, wavelet-based and NSCT-based algorithms in terms of human vision and the object criteria. The textures are more distinct, the contrast is more preferable, and detail variance (DV) of proposed method is much higher than the other methods, while the background variance (BV) is almost equal.