DSTnet: a new discrete shearlet transform-based CNN model for image denoising

Abstract

Due to the superior performance and fast running speed, deep learning methods have been widely employed in image processing fields. However, most deep learning-based denoising methods require a specific noise level to train their models, and denoising models are built to remove specific levels of noise, which lacks the flexibility to deal with spatially variant noise. In this paper, we present a discrete shearlet transform (DST)-based denoising convolutional network (DSTnet). The proposed method first decomposes an image by DST and gets several subband images, then these subband images are used as input samples into the convolutional neural networks (CNNs) blocks. The proposed method has a good compromise between denoising performance and computation time. The DSTnet not only has a good efficiency in noise removing and detail preservation but also has the ability to handle a wide range of noise levels, which are suitable for real image denoising.