The Research And Application Of The Information Extraction Algorithm In FTIR Multi-spectral Microscopic Images

Fourier transform infrared (FTIR) micro-spectroscopic imaging is a new and potential tool for subtle changes detection, which is developing from FTIR microscopy. The outstanding advantages of FTIR micro-spectroscopic method are little interferer, non-destructive and reproducible, especially the original and direct analysis to fresh tissues. Moreover, this technique provides the opportunity to visualize the distribution of chemical ingredients among the mixture. In most conventional methods, identification of subtly chemical change remains a time consuming and destructive technique, which is directly related to the experience and capacity of scientific researchers. Over the past decades, Fourier transform infrared spectroscopy has got a considerable amount of attention on food and drug testing, medical diagnosis, chemical composition analysis and so on. Our present studies shall attempt to explore strategies that may allow detecting and extracting particular distribution of chemical composition. The strategies are named information extraction, and are used to food testing and medical diagnosis.This thesis mainly studies on the the information extraction strategies which is based on Principal Component Analysis (PCA) and Least Square Support Vector Machine (LS-SVM). As the earlier preparation of information extraction, firstly, this paper briefly introduces the basical theories of FTIR multi-microspectroscopic imaging and sample preparation, and then subtly describes the methods of data pre-processing under different conditions. As is shown in experimental results which related with rabbit artery sample and wheat kernel sample, this method achieves higher data quality.Secondly, the fundamental algorithms of Principal Component Analysis are deeply discussed, and the new extractional methods based on Lambert-Bill model is proposed, which is named Lamber Beer PCA (LPCA). Target features are selected by comparing the spectral similarities of loading vectors and standard sample spectrums. As is shown in theoretical analysis, the new method has advantages in mining data structures of chemical information. According to the inherent flaws of the global algorithm in dealing with large samples, new improved algorithm (Improved LPCA) is proposed which combined with data sources division and peak judgment to improve performance of algorithm. Thirdly, some mainstream classification algorithms are briefly introduced, and the basic theories of the Least Squares Support Vector Machine (LS-SVM) are discribed, including the methods of data training and the parameter selection. Human-computer interaction methods are proposed for improveing flexibility and accuracy. The experimental results indicate that the new classification algorithm is effective.Finally, as two groups of experimental results shown, compared with the exsiting methods, the new methods can achieve higher accuracy and efficiency. In a word, it's evident that new method has the advantages in infrared multi-microspectral images.