Rethinking the CSC Model for Natural Images

Sparse representation with respect to an overcomplete dictionary is oftenused when regularizing inverse problems in signal and image processing. Inrecent years, the Convolutional Sparse Coding (CSC) model, in which thedictionary consists of shift-invariant filters, has gained renewed interest.While this model has been successfully used in some image processing problems,it still falls behind traditional patch-based methods on simple tasks such asdenoising. In this work we provide new insights regarding the CSC model and itscapability to represent natural images, and suggest a Bayesian connectionbetween this model and its patch-based ancestor. Armed with these observations,we suggest a novel feed-forward network that follows an MMSE approximationprocess to the CSC model, using strided convolutions. The performance of thissupervised architecture is shown to be on par with state of the art methodswhile using much fewer parameters.