Regression and sparse regression methods for viscosity estimation of acid milk from it´s SLS features

Statistical solutions find wide spread use in food and medicine quality control. We investigate the effect of different regression and sparse regression methods for a viscosity estimation problem using the spectro-temporal features from new Sub-Surface Laser Scattering (SLS) vision system. From this investigation, we propose the optimal solution for regression estimation in case of noisy and inconsistent optical measurements, which is the case in many practical measurement systems. The principal component regression (PLS), partial least squares (PCR) and least angle regression (LAR) methods are compared with sparse LAR, lasso and Elastic Net (EN) sparse regression methods. Due to the inconsistent measurement condition, Locally Weighted Scatter plot Smoothing (Loess) has been employed to alleviate the undesired variation in the estimated viscosity. The experimental results of applying different methods show that, the sparse regression lasso outperforms other methods. In addition, the use of local smoothing has improved the results considerably for all regression methods. Due to the sparsity of lasso, this result would assist to design a simpler vision system with less spectral bands.