A Novel Supervised Distributed Dictionary Learning based on Learned K-SVD for Image Denoising

With limited computational resources and storage space on SoCs, deploying large deep-learning networks (DNNs) is challenging. However, dictionary learning (DicL) has lower complexity and storage space requirements while improving the interpretability of the model. In addition, deep unfolding techniques can construct high-performance end-to-end networks based on DicL. Therefore, in this paper, we apply Deep KSVD (LKSVD), a deep unfolding network based on the classical K-SVD algorithm, with a distributed framework and propose a distributed dictionary learning (DDL) method called DDL-LKSVD. We experimentally validate DDL-LKSVD on the classical and fundamental image denoising problem, and the experimental results show that the average PSNR values we achieve in the DDL-LKSVD proposed in this paper on the Set12 dataset with the noise level 25 are 2.58 dB, 2.74 dB, 0.29 dB, and 0.03 dB higher than those of OMP, ISTA, K-SVD, and LKSVD, respectively.