Image Denoising: An Approach Based on Wavelet Neural Network and Improved Median Filtering

Abstract

Wavelet neural network (WNN) is introduced into the field of digital image denoising due to the excellent local feature and adaptive ability. The procedure of denoising can be looked as an approximating procedure from noisy image to original image. The better WNN has approximation performance, the better denoising performance. Researches have shown that WNN does well in the approximation of nonlinear function and consequently it can be employed in image denoising. In our denoising approach, feature extracting is performed with the help of an improved median filtering, and feature values normalized by exponential function are input into the WNN trained by the classic but efficient gradient descent method (GDM). The feature value is sensitive to the pepper noise in dark background and salt noise in normal background. The experimental results showed that the proposed approach was superior to traditional median filtering in the ability of preserving fine details and excellent fidelity and also indicated the efficiency of the proposed approach in the case of high intensity noise