Adaptive dictionary learning algorithms jointly optimize sparse representations and dictionary atoms to enhance performance. However, this alternating minimization process is complex, and the two optimizations often remain independent of each other, limiting effectiveness. In this work, we propose a new adaptive dictionary learning framework, termed Projected Dictionary Learning (PDL). The key idea of PDL is to update only the important dictionary atoms (IDA) by leveraging the projections of effective input vectors (EIV)—that is, input samples whose sparse representations share activations at the same atom index. This projection-driven update strategy enables PDL to better capture structural information during training. Experimental results demonstrate that PDL provides superior image denoising performance compared to both Orthogonal Matching Pursuit (OMP) and K-SVD. Unlike OMP, which relies on a fixed incoherent dictionary, and K-SVD, whose performance is strongly dependent on the choice of the initial dictionary, PDL consistently achieves robust results across a variety of initial dictionaries, standard test images, and a wide range of additive noise levels. These results highlight the stability, efficiency, and effectiveness of PDL as a reliable dictionary learning method for sparse coding–based image denoising.
Hoori et al. (Thu,) studied this question.