Research On Terahertz Spectral Identification With Small Samples

The terahertz spectrum of a substance has a unique "fingerprint spectrum" characteristic,so it can be used to identify the substance.With the development of artificial intelligence technology,machine learning and deep learning algorithms have been used more and more widely in the field of terahertz spectrum recognition.Thanks to the integration of powerful deep learning technology and large-scale labeled training datasets,large-scale target recognition has reached a high level of performance.However,in practical applications,due to the influence of experimental equipment,experimental conditions,and experimental environment,the terahertz spectral data we obtain are not always large-scale,and even the terahertz spectral data of materials within a certain sampling frequency are missing cases,which cannot satisfy the large amount of data required by deep learning algorithms.How to effectively use these small sample data to identify substances in the field of terahertz spectrum identification is a big problem.To solve this problem,this paper explores a terahertz spectrum recognition method based on transfer learning.First,SG filtering and cubic spline interpolation were used to preprocess the terahertz transmission spectrum data of Maltotriose,Maloheptaose,Maltotetraose,Maltopentaose,and Malthexaose within 0.9-6Thz.the terahertz transmission spectra of Maltotriose and Maloheptaose at 0.9-6thz were used the source domain dataset,and the terahertz transmission spectrum of Maltotetraose,Maltoopentaose,and Malthexaose at 0.9-6thz were used as the target domain dataset.Then we trains a Convolutional neural network(CNN)to extract the spectral features of the source dataset and extract them.The features of the source domain dataset are transferred to the target domain dataset,and finally the spectral features obtained by the fusion are input to the integrated classifier for spectrum recognition.The experimental results show that the terahertz spectrum recognition method based on feature transfer can effectively solve the overfitting problem and improve the accuracy of terahertz spectrum recognition in the case of small samples.The terahertz spectrum recognition method based on feature migration also has certain limitations.When the correlation between the source domain data set and the target domain data set is low,the migration effect will be greatly reduced,and even negative migration will occur.Besides,the method of enhancing and identifying terahertz spectrum data of small samples is also studied in this paper.This method can not only repair the defected terahertz spectral data,but also generate data to enhance the diversity of spectral data,so as to better identify substances.First we use the same data processing method to preprocess the spectral data of ten substances(Anthraquinone,Benomyl,Carbazole,Mannose,Riboflavin,Malthexaose,Maltoheptaose,Maltoopentaose,Maltotetraose,Maltotriose),and then generate adversarial networks(GAN)will be used to repair the missing terahertz spectrum data.In addition,GAN can also generate simulation data with real terahertz spectrum data distribution.Finally,the repaired data,generated data,and real spectral data are used as training samples to train the deep neural network to obtain the substance recognition results.The experimental results show that GAN can effectively repair the missing terahertz spectrum data,and the generated terahertz spectrum data effectively simulates the overall characteristics of real terahertz spectrum data,adds terahertz spectrum data samples,and greatly improves spectral recognition Precision.