Rethinking the CSC Model for Natural Images

NeurIPS 2019  ·  Dror Simon, Michael Elad ·

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.

PDF Abstract NeurIPS 2019 PDF NeurIPS 2019 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Color Image Denoising BSD68 sigma15 CSCNet PSNR 33.83 # 3
Color Image Denoising BSD68 sigma25 CSCNet PSNR 31.18 # 2
Color Image Denoising BSD68 sigma75 CSCNet PSNR 26.32 # 1
Color Image Denoising CBSD68 sigma50 CSCNet PSNR 28.00 # 10

Methods


No methods listed for this paper. Add relevant methods here