Simultaneous regularized sparse approximation for wood wastes NIR spectra features selection

This paper presents a new technique of simultaneous sparse approximation incorporating a regularity constraint along the coefficients matrix rows. This approach is decomposed in two steps: first a sparse representation of the coefficients matrix is obtained using a simultaneous greedy method. Then, a ℓ1 penalty regularization on the derivative of nonzero coefficients enforces a piecewise constant variation along the rows of the solution. The regularization problem is solved efficiently using the ADMM (Alternate Direction Method of Multipliers) optimization method. The approach is applied on near-infrared spectrometry dataset of wood wastes. This allows to select among the 1647 wavelengths of the spectra those suitable for classification. The experimental tests validate the advantages of regularization in terms of classification rates.