DCTNet: deep shrinkage denoising via DCT filterbanks

Shrinkage algorithms are well-studied, simple yet efficient transform domain denoisers. Two factors greatly affect their performance, namely the types of signal transform and shrinkage function which are used. The purpose of this study is to develop novel deep learning-based variants for transform domain shrinkage approaches. In particular, the Discrete Cosine Transform (DCT) will be considered as the sparsifying transform utilized in conjunction with deep neural networks. There has been comparatively few studies for the amalgamation of the DCT and deep learning compared to other transforms such as Discrete Wavelet Transform (DWT). Main reason for this is the fact that both global and block treatments of the DCT do not provide feature maps (that is subband images) suitable for processing by deep convolutional neural networks (CNNs). On the other hand, researchers have regularly modeled learnable shrinkage functions that are tuned to satisfy properties such as symmetry and monotonicity while restricting denoiser’s performance. In this paper, we propose a novel DCT-based deep denoising algorithm which consists of three blocks: an original DCT block, a deep shrinkage block, and an inverse DCT block. DCT blocks use 2D DCT basis kernels as mapping filters. The resulting transform is called DCT filterbanks (DCT FB) transform. Proposed DCT FB blocks facilitate the effective production of DCT subband images suitable for processing by CNNs. Instead of analytic shrinkage step, the shrinkage operation is parameterized with deep learning layers and is called as shrinkage block. The proposed DCT domain deep shrinkage network, termed as DCTNet, is trained in a supervised manner and provides an effective and improved hybrid of classical patchwise shrinkage algorithms with deep learning. Our experimental results indicate that the proposed method surpasses model-based and deep CNN-based denoisers.