NSTBNet: Toward a nonsubsampled shearlet transform for broad convolutional neural network image denoising

Abstract

Deep convolutional neural networks (CNNs) have achieved huge success in the image denoising fields. However, there still have three drawbacks to be overcome: firstly, it is a big cost to train a deeper CNN model for better denoising performance, secondly, most models often lack the flexibility to remove spatially variant additive Gaussian noise (AWGN), finally, it is difficult to achieve a good trade-off between denoising performance and details preservation. In this paper, we propose a novel network integrated with nonsubsampled shearlet transform (NSST) and a broad convolutional neural network, namely NSTBNet. The proposed NSTBNet has the following desirable properties: (1) a single model has the ability to deal with different noise levels and spatially variant AWGN; (2) the model combines the two networks make it wider to improve denoising performance without increasing too much computational cost; (3) the NSST and inverse NSST are employed to acquire more texture and to avoid gridding effect. Extensive simulation experiments on AWGN and realistic noise show that NSTBNet has state-of-the-art denoising performance, making it a high application value.