Connectogram - A graph-based time dependent representation for sounds

Abstract

The proposed method contributes the time-series classification literature with a novel time-convexity based representation, which extends the current graph conversion approaches by introducing the time dimension, also introducing a colorful graph-generator approach. The representation capability of connectograms is tested in comparison with mel-spectrograms (mels) and MFCCs for an environmental sound classification task, as input to state-of-art transfer learning models. Results indicate that connectograms cannot compete with the best-performer mel-spectrogram representations in standalone format, however they significantly improve their classification performance in case they are combined as single layers of hybrid RGB representations. A combination of [mels + mels + connectogram] outperforms either sole representations or their combinations by 2-3%, with 96.46% classification accuracy for ResNet50 classifier model.(c) 2022 Elsevier Ltd. All rights reserved.