A Convolutional Dictionary Learning based l1 Norm Error with Smoothed l0 Norm Regression
Abstract: General sparse representation schemes for image signals divide given images into non-overlapped patches, and discuss dictionaries and/or sparse coefficients for the patches. A convolutional sparse representation provides an alternative approach, in which an entire image is represented by the sum of convolutions with dictionary filters. This work tackles designing of the dictionary filters with the l 1 norm error criterion in the data fidelity term to improve robustness against outliers. This paper also applies smoothed l 0 norm to estimate sparseness of coefficient vectors.
Loading