Data Augmentation for Mixed Spectral Signatures Coupled with Convolutional Neural Networks

Abstract

Advances in photo-thermal infrared imaging spectroscopy technologies have come a long way. These advances include the development of quantum cascade laser (QCL) which greatly improves the measurement of infrared absorption. Despite the progress made, a few fundamental limitations of infrared spectroscopy have prevented its processing in deep learning models. Direct application of convolutional neural networks becomes difficult, because of the high dimensionality, and low number of samples. To overcome these limitations, in this paper, we explore the applicability of data augmentation techniques to photo-thermal infrared imaging data. First, we establish a baseline using the raw data and a convolutional neural network (CNN). Second, we apply principal component analysis (PCA) and proceed to applying CNN. Third, we employ different data augmentation techniques before applying CNN. After that, we estimate the performance of our models by using k-fold cross-validation and calculating the confusion matrix and classification accuracy. Our GDAC+CNN model achieved an average of 99% accuracy.