New Preprocessing Methods of Raman Spectra and their Evaluation by Wu Yingwen Thesis Supervisor: Assistant Professor Long CHEN Department of Computer and Information Science University of Macau Raman spectroscopy, as a non-invasive optical technique, can provide characteristic information of the molecule structure, which is widely used in many fields such as chemistry, materials science, biology, medicine and so on. However, Raman spectra include some interferences other than the useful spectral information, which have a negative effect on the analysis. Therefore, it is crucial to conduct appropriate preprocessing procedures for the spectra data before analysis. Here, we present a comprehensive investigation of these preprocessing procedures, which include denoising, background correction, and normalization. Among them, the background correction approach plays a key role particularly. Multiple spectra, generated by static scanning every few seconds of sample, have a close relationship between each other in the same set of simultaneously collected data. They are consist of duplicate signals with similar noise and background information. Inspired by this, we take advantage of the inherent property of this kind of closely related data to improve the background correction performance. Here, based on the weighted penalized least squares methods of estimating the background, we propose two new schemes of weight calculation. Both are formed by averaging the weights of all the spectra in the set. To some extent, the calculation process of averaging weight can suppress the noises and spectral variations in the data. As a result, the estimated background could be more accurate with the collaborative consideration of weight. We also take two existing penalized least square based methods as an example to combine them with our collaborative processing methods. And we think other penalized least squares based background correction methods could also be improved in the same way. Besides, in order to better verify background correction methods, we design a series of evaluation methods both for simulated data and real Raman spectra data. Furthermore, we want to extend the collaborative processing methods to deal with the spectra collected from Raman mapping, which also have closely internal correlation in a same set. In order to remove the irrelevant spectra, we firstly use the kernel-based fuzzy C-means to cluster the data. We then extract the set with obvious characteristic peak information and carry out the following processing works as multiple spectra. Above all, we have successfully proposed two new preprocessing methods of background correction for Raman spectra. And a set of systematic evaluation methods is also introduced to test the performance of new methods. Compared with traditional single point spectrum based preprocessing methods, our improved methods achieve better performance for the multiple spectra both on simulated data and on real Raman spectra data. We hope this collaborative processing idea could be extended by some other aggregation operators when confronted with the related spectra such as multiple spectra and Raman mapping spectra.