Spectral pattern recognition and application

Abstract

Spectroscopy coupled with pattern recognition methods, spectral pattern recognition, is used in many areas of science and engineering to analyse the constituent of matter at the atomic and molecular levels by exploring the interaction between constituent of matter and radiated energy. It provides a fast and nondestructive approach for chemical detection, which is highly efficient compared to traditional techniques such as wet chemistry and chromatography. A common challenge for spectral pattern recognition is how to use mathematical, statistical and machine learning methods to effectively perform quantitative or qualitative analysis on noisy and variable spectral data generated by spectroscopic systems, especially low-cost ones. In this thesis, we seek to address this challenge; in particular, we focus on the specific issue of analysing spectral data with complex structures (high-dimensionality and high-collinearity) to extract useful fingerprint information from noisy field data and assist low-cost spectroscopic systems to achieve the required level of performance for food authentication.

Partial least squares discriminant analysis (PLS-DA) is one of the most frequently used multivariate analysis methods for spectral data analysis, which can handle high-dimensionality and high-collinearity in spectral data. We develop a local modelling based variant of PLS-DA in order to improve the classification performance of PLS-DA on high-dimensional, nonlinear and multimodal data. The resulting local PLS-DA models have simple yet discriminant structures which are approximately linearly-separable by using a small number of latent variables. We combine collaborative representation-based classifier (CRC) with weighted PLS regression in order to provide a simple and intuitive interpretation of how each variable contributes to the classification decision of a query. We evaluate the above methods through comparison with baseline methods on public machine learning and spectral datasets through a series of experiments. Moreover, we develop a low-cost sensor system using diffraction grating, torch and camera, which can produce nonstandard spectral data via computer vision techniques. With the aid of state of the art classifiers, this sensor system achieves comparable results to portable spectroscopy for organic apple authentication, demonstrating the potential of our work in the provision of a viable solution to empower consumers in food authentication.

Date of Award	Apr 2019
Original language	English
Supervisor	H. Wang (Supervisor), Paul Maguire (Supervisor) & Omar Nibouche (Supervisor)