Concatenation-based CNN feature fusion achieved 98.45% accuracy and 97.94% Kappa for CWT-based EEG motor imagery classification, outperforming individual CNN extractors and summation fusion.
Does CNN feature fusion improve CWT-based EEG motor imagery classification compared to individual CNN extractors?
Concatenation-based CNN feature fusion significantly improves EEG motor imagery classification accuracy over individual CNN extractors.
ABSTRACT This letter investigates the effect of convolutional neural network (CNN) feature fusion on electroencephalography (EEG)‐based motor imagery classification. EEG trials are converted into time–frequency scalogram images using the continuous wavelet transform (CWT). Three pre‐trained CNNs, namely GoogleNet, AlexNet and ResNet50, are used as feature extractors. Their individual performance is compared with two fusion strategies: summation and concatenation. The main contribution of this letter is to isolate the CNN feature‐fusion stage and examine how the choice of fusion mechanism affects CWT‐based MI‐EEG classification. Experiments on the BCI Competition IV Dataset 2a show that feature fusion improves classification performance over individual CNN extractors. Concatenation‐based fusion achieves the best result, reaching 98.45% accuracy and 97.94% Kappa. The results indicate that preserving complementary CNN features is more effective than compressing them through summation for EEG motor imagery classification. Unlike complete EEG‐MI decoding frameworks, this letter focuses specifically on isolating and analysing the effect of the CNN feature‐fusion mechanism under a fixed CWT‐based evaluation setting.
Jihad et al. (Thu,) conducted a other in EEG motor imagery classification. CNN feature fusion (concatenation) vs. Individual CNN extractors and summation fusion was evaluated on Classification accuracy and Kappa. Concatenation-based CNN feature fusion achieved 98.45% accuracy and 97.94% Kappa for CWT-based EEG motor imagery classification, outperforming individual CNN extractors and summation fusion.