Robust neural network threshold determination for wavelet shrinkage in images

The discrete wavelet transform (DWT) has been established as an effective tool in denoising images. Various studies have developed statistical models for denoising signals in the wavelet domain. In these techniques, the amount of noise is estimated from the detail coefficients of the transform. However, in images rich in textures, this estimate does not accurately reflect the noise levels of the image. In this paper, we introduce a robust method of noise and signal estimation using directional characteristics of an image. A feed-forward neural network is utilized to establish the relationship between the new estimators and the optimal soft threshold. Testing results show equivalent performance to traditional thresholding algorithms in most images. In highly detailed images, the proposed network shows significant improvement in denoising.