WGLFNets: Wavelet-based global–local filtering networks for image denoising with structure preservation

Abstract

Efficiency and denoising performance are two important indexes to objectively evaluate an image denoising method. Traditional model-based denoising methods usually pursue pleasing performance at the cost of highly computational complexity, while the popular deep learning-based methods perform well on efficiency improvement but still fail in rich structure preservation. To cope with these problems, this paper proposed a novel wavelet-based deep denoising methods, called as wavelet-based global–local filtering networks (WGLFNets). The WGLFNets includes the following three key points: First, an noisy image is decomposed into a base image and three-directional detail images by employing discrete wavelet transform (DWT). Second, the base component is denoised by a global denoising network while the detail components are recovered by a local detail restoration network. Finally, all processed components are utilized to construct the high-quality image by the inverse DWT (IDWT). Experimental results demonstrate that the WGLFNets is effective and is superior to the state-of-the-arts.