[11a-W810-1~10] 31.1 フォーカストセッション「AIエレクトロニクス」

▼ [11a-W810-9] Generative adversarial network for robust Raman spectra identification

〇(M2)HSUNWEN FANG¹、CHUNHWAY HSUEH¹ (1.National Taiwan Univ.)

キーワード：adversarial network, artificial classifier, Raman spectra

Generative adversarial network (GAN) is a reinforcementmachine learning technique where two machines contest inlearning abilities. This algorithm was often realized byconstructing two neural networks (NN) where one evaluatesdata (discriminator) and the other generates confusing data(generator) to deceive the other one. Recently, the effects ofGAN on discriminators have incurred some attentions due toits efficacy to improve the network robustness.1Raman spectroscopy is a highly promising technique forexplosives detection due to its label-free and stand-offnature.2 However, as few as 1 in 106–108 photons scatters asRaman signal, which makes Raman detections susceptible tofluorescence and environmental noises. This problemdeteriorates in situations of distant detection, which isindispensable in Raman detection of explosives.In this work, a semi-supervised GAN classifier wasconstructed to classify the Raman spectra of the mostprevalent explosive, TNT, and its precursor, 24-DNT. Withhighly resembling Raman spectra accompanied by severenoises, these chemicals were often hard to be distinguishedthrough Raman detections. Herein, the GAN classifier withbetter network robustness was adopted which outperformedits supervised counterparts due to superior network robustness.