Rethinking the CSC Model for Natural Images

TL;DR

This paper offers new insights into the convolutional sparse coding (CSC) model for natural images, connecting it to patch-based methods and proposing a novel efficient neural network that achieves competitive results.

Contribution

It introduces a Bayesian perspective on CSC, links it to patch-based models, and develops a new feed-forward network using strided convolutions for improved efficiency.

Findings

The proposed network matches state-of-the-art performance.

The model uses significantly fewer parameters.

Provides theoretical insights connecting CSC and patch-based methods.

Abstract

Sparse representation with respect to an overcomplete dictionary is often used when regularizing inverse problems in signal and image processing. In recent years, the Convolutional Sparse Coding (CSC) model, in which the dictionary 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 as denoising. In this work we provide new insights regarding the CSC model and its capability to represent natural images, and suggest a Bayesian connection between this model and its patch-based ancestor. Armed with these observations, we suggest a novel feed-forward network that follows an MMSE approximation process to the CSC model, using strided convolutions. The performance of this supervised architecture is shown to be on par with state of the art methods while using much fewer parameters.