Multivariate Analysis Techniques for Optimal Vision System Design

The present thesis considers optimization of the spectral vision systems used for quality inspection of food items. The relationship between food quality, vision based techniques and spectral signature are described. The vision instruments for food analysis as well as datasets of the food items used in this thesis are described. The methodological strategies are outlined including sparse regression and pre-processing based on feature selection and extraction methods, supervised versus unsupervised analysis and linear versus non-linear approaches.

One supervised feature selection algorithm based on the existing sparse regression methods (EN and lasso) and one unsupervised feature selection strategy based on the local maxima of the spectral 1D/2D signals of food items are proposed. In addition, two novel feature extraction and selection strategies are introduced; sparse supervised PCA (SSPCA) and DCT based characterization of the spectral diffused reflectance images for wavelength selection and discrimination.

These methods together with some other state-of-the-art statistical and mathematical analysis techniques are applied on datasets of different food items; meat, diaries, fruits and vegetables. These datasets are acquired using three different vision systems; a spectral imaging device called VideometerLab, spectroscopy, and diffused reflectance imaging systems called Static Light Scattering (SLS).

These analyses result in signifcant reduction in the number of required wavelengths and simplifcation of the design of practical vision systems.