Sparse Bayesian learning approach for baseline correction

Abstract Spectral techniques in analytical chemistry are often affected by baselines in practical implementation. Without baseline correction, the accuracy of the qualitative/quantitative analytical results may degrade substantially. Sparse representation has been applied to baseline correction recently, which can provide state-of-the-art performance. However, it suffers from possible performance degradation when realized using l 1 -norm approximation. To significantly improve the performance for baseline correction, a sparse Bayesian learning (SBL) framework for joint pure spectrum fitting and baseline correction is presented in this work. Since the SBL framework provides high flexibility to tackle the minimum l 0 -norm problem instead of the l 1 -norm approximation, it is possible to yield higher baseline correction accuracy. Moreover, the proposed method has an inherent learning capability, so no additional regularization terms are required. Note that the sparse representation performance would degrade if the grid points used in dictionary matrix are not sufficiently dense. Therefore, we further consider grid points as adjustable parameters and then adopt a grid refinement technique to handle the off-grid gap. Results on both simulated and real datasets reveal substantial performance improvement of the proposed SBL-based method over the existing schemes on baseline correction.