Research On The Preprocessing And Recognition Algorithms Of Plasmon-enhanced Raman Spectra

Substances detection plays an important role in our life and work. Raman spectroscopy technique is getting more and more attentions in the field of rapid detection because of its features of rapidness and non-destruction. The existing analysis algorithms of Raman spectra have many limitations. They always focus on unique instruments or just work on single bath substances. And the operators of these unique instruments must be professional in chemistry. We first propose two preprocessing algorithms of Raman spectra which automatically and efficiently remove fluorescence background without manual intervention. And they are both versatile. One algorithm is based on locally adaptive polynomial fitting and the other is based on the Gaussian hypothesis. We also propose three solutions based on machine learning to recognize and classify Raman spectra according to different demands.(1) As for the demand of multi-classification, logistic regression is used to categorize a Raman spectra into a set of classes. And the method based on min-max signal adaptive zooming can further improve the accuracy of classification.(2) Hypergraph learning algorithm can solve the problem due to the lack of labeled training samples, in which unlabeled samples of Raman spectra are also used for training.(3) With the benefits of stacked denoising autoencoders, high level feature representations are abstracted from original Raman spectra. Then we can recognize and classify Raman spectra without prior-knowledge. Several factors that affect the performance of stacked denoising autoencoders are also discussed.