Extending CNN Classification Capabilities Using a Novel Feature to Image Transformation (FIT) Algorithm

Abstract

In this work, we developed a novel approach with two main components to process raw time-series and other data forms as images. This includes a feature extraction component that returns 18 Frequency and Amplitude based Series Timed (FAST18) features for each raw input signal. The second component is the Feature to Image Transformation (FIT) algorithm which generates uniquely coded image representations of any numerical feature sets to be fed to Convolutional Neural Networks (CNNs). The study used two datasets: 1) behavioral biometrics dataset in the form of time-series signals and 2) EEG eye-tracker dataset in the form of numerical features. In earlier work, we used FAST18 to extract features from the first dataset. Different classifiers were used and Deep Neural Network (DNN) was the best. In this work, we used FIT on the same features and invoked CNN which scored 96% accuracy surpassing the best DNN results. For the second dataset, the FIT with CNN significantly outperformed DNN scoring ~90% compared to ~60%. An ablation study was performed to test noise effects on classification and the results show high tolerance to large noise. Possible extensions include time-series classification, medical signals, and physics experiments where classification is complex and critical.