The study of electroencephalogram (EEG) signals is crucial for understanding brain function and has extensive applications in clinical diagnosis, neuroscience, and brain–computer interface technology. This paper addresses the challenge of recognizing motor imagery EEG signals with few channels, which is essential for portable and real-time applications. A novel framework is proposed that applies a continuous wavelet transform to convert time-domain EEG signals into two-dimensional time-frequency representations. These images are then concatenated into channel-dependent multilayer EEG time-frequency representations (CDML-EEG-TFR), incorporating multidimensional information of time, frequency, and channels, allowing for a more comprehensive and enriched brain representation under the constraint of few channels. By adopting a deep convolutional neural network with EfficientNet as the backbone and utilizing pre-trained weights from natural image datasets for transfer learning, the framework can simultaneously learn temporal, spatial, and channel features embedded in the CDML-EEG-TFR. Moreover, the transfer learning strategy effectively addresses the issue of data sparsity in the context of a few channels. Our approach enhances the classification accuracy of motor imagery EEG signals in few-channel scenarios. Experimental results on the BCI Competition IV 2b dataset show a significant improvement in classification accuracy, reaching 80.21%. This study highlights the potential of CDML-EEG-TFR and the EfficientNet-based transfer learning strategy in few-channel EEG signal classification, laying a foundation for practical applications and further research in medical and sports fields.

Electroencephalogram (EEG) signals hold significant research value in brain function decoding, disease diagnosis, and brain–computer interfaces (BCIs). Few-channel EEG recording devices feature superior portability, simple operation, and facilitated real-time monitoring implementation. However, few-channel motor imagery (MI) EEG signals inherently suffer from data scarcity and limited spatial discriminative information, which pose critical challenges, including insufficient feature extraction and poor robustness in classification tasks. To address these issues, this paper presents EEG–ShuffleFormer, a hybrid network that integrates two complementary views of EEG signals: time–frequency representations obtained via continuous wavelet transform and the original raw signal representations. A lightweight ShuffleNet backbone extracts local features, followed by a Transformer encoder that models long-range temporal dependencies. Evaluated on the BCI Competition IV Dataset 2b, the proposed method achieves an average classification accuracy of 82.23%, with a substantial improvement on challenging subjects compared to the closest baseline method. Compared with existing methods, the proposed multi-view fusion strategy raises the performance floor while maintaining high accuracy on typical subjects, demonstrating its potential to enhance robustness for different subjects in few-channel scenarios.