Machine learning for pattern and waveform recognitions in terahertz image data

Several machine learning (ML) techniques were tested for the feasibility of performing automated pattern and waveform recognitions of terahertz time-domain spectroscopy datasets. Out of all the ML techniques under test, it was observed that random forest statistical algorithm works well with the THz datasets in both the frequency and time domains. With such ML algorithm, a classifier can be created with less than 1% out-of-bag error for segmentation of rusted and non-rusted sample regions of the image datasets in frequency domain. The degree of linear correlation between the rusted area percentage and the image spatial resolution with terahertz frequency can be used as an additional cross-validation criteria for the evaluation of classifier quality. However, for different rust staging measured datasets, a standardized procedure of image pre-processing is necessary to create/apply a single classifier and its usage is only limited to 1 +/- 0.2 THz. Moreover, random forest is practically the best choice among the several popular ML techniques under test for waveform recognition of time-domain data in terms of classification accuracy and timing. Our results demonstrate the usefulness of random forest and several other machine learning algorithms for terahertz hyperspectral pattern recognition.

Automated summary

Several machine learning techniques were tested for automated pattern and waveform recognition of terahertz time-domain spectroscopy datasets, with random forest algorithm showing good performance. Linear correlation and additional cross-validation criteria can be used to evaluate classifier quality, requiring a standardized image pre-processing procedure for different rust staging datasets. Random forest is practically the best choice for waveform recognition in terms of classification accuracy and timing.