Thresholding neural network (TNN) with smooth sigmoid based shrinkage (SSBS) function for image de-noising

Abstract

In this paper we proposed a new method for noise removal in wavelet domain. In this method we developed a thresholding neural network (TNN) by using a new type of smooth nonlinear thresholding function as its activation function. With respect to this function gradient based adaptive learning algorithm becomes more efficient in finding the optimal threshold to obtain least mean square (LMS) or minimum mean square error (MMSE). Experimental results shows that TNN with adaptive learning algorithm (TNN based nonlinear adaptive filtering) outperforms some other alternative methods in image de-noising in terms of obtaining higher peak signal to noise ratio (PSNR) and visual quality. The proposed method achieves up to 3.48 dB improvement over the state-of-the-art for de-noising Cameraman image.