Presentation Information
[A-10-26]Consideration of Deep Dictionary Learning with Fast Iterative Shrinkage-Thresholding Algorithm Unfolding
◎Ataru Shakagori1, Michiharu Maeda1 (1. Graduate School of Engineering, Fukuoka Institute of Technology)
Keywords:
dictionary learning,deep learning,fast iterative shrinkage-thresholding algorithm,mutual coherence constraint,image denoising
This paper investigates the effects of network unfolding and structural constraints on image denoising performance, with a specific focus on the deep K-singular value decomposition (K-SVD) network. Conventional approaches predominantly rely on the iterative shrinkage-thresholding algorithm (ISTA), leaving the potential of accelerated unfolding methods largely unexplored. The objective of this work is to evaluate whether a fast ISTA (FISTA) unfolding combined with a mutual coherence constraint can serve as a more effective and structurally optimized framework for dictionary learning. We specifically investigate deep K-SVD networks unfolded via LFISTA, integrating a penalty term that minimizes the off-diagonal elements of the Gram matrix. By evaluating representation errors and atom coherence during learning, we aim to clarify how the structural optimization affects dictionary's structural adaptability and final reconstruction quality. Experimental results demonstrate that our LFKSVD manner decreases the average coherence of the dictionary and improves the overall peak signal-to-noise ratio (PSNR) compared to the conventional LKSVD network.