A Novel Curvelet Thresholding Function for Additive Gaussian Noise Removal

Improving quality of noisy images has been an active area of research in many years. It has been shown that removing additive Gaussian noise by nonlinear methods such as Wavelet denoising and Curvelet denoising had better results than classic approaches. However estimation of threshold and selection of thresholding function are still challenging tasks. In this paper, a new thresholding function is proposed for Curvelet thresholding and thresholding neural network is extended to use Curvelet coefficients instead of wavelet coefficients. This function is continues and has higher order derivation. Therefore it is suitable for gradient decent learning methods such as thresholding neural network (TNN). This function is used by the TNN and threshold values for Curvelet sub-bands are estimated according to least mean square (LMS) algorithm. The experimental results show improvement in noise reduction from images based on visual assessments and PSNR comparing with well-known thresholding functions.