Abstract
Visual displays of data are the main means of communicating numerical analytics in an intuitive and efficient way. Graphical concepts, such as colors, shapes, and sizes enable embedding of a broad range of information into visualisations, and hence reduce the dimensionality of data. Furthermore, with the advancements in Convolutional Neural Networks (CNN), data visualisations have found their way in deep learning studies for segmentation, prediction, and pattern recognition purposes. To this extent, a signal can be initially transformed into a comprehensive visualisation such as time-frequency representation (TF), and be used as the input layer in the CNN model. Color, as a key feature for TF representation, plays an important role in delivering the information. In this study, we investigate the use of TF representations and colormaps in CNN models. We show that use of an inappropriate colormap for TF representation of a signal can affect the accuracy of a CNN model, even up to 24% for a specific task. This is mainly due to the performance of colormaps which makes them susceptible to the stochastic nature of CNN models. We, further, introduce a simple methodology to examine the effectiveness of colormaps in CNN feature learning and extraction in regard to the signal TF representation.
Original language | English |
---|---|
Article number | 121889 |
Pages (from-to) | 1-15 |
Number of pages | 15 |
Journal | Information Sciences |
Volume | 702 |
Early online date | 27 Jan 2025 |
DOIs | |
Publication status | Published online - 27 Jan 2025 |
Bibliographical note
Publisher Copyright:© 2025
Data Access Statement
All the raw datasets used in this study can be accessed online through their associated references. Two sample codes for data preparation and CNN training is added to GitHub https://github.com/pardis-pb/Colormap_CNN.git.Keywords
- Colormap
- Time-Frequency
- Non-stationary signals
- Pattern Recognition
- Convolutional Neural Network
- Deep Learning
- Time-frequency (TF)
- Convolutional neural network (CNN)
- Pattern recognition
- Deep learning (DL)