Objective This study addresses the limitations of deep learning-based methods in recognizing motor imagery electroencephalography EEG signals which primarily focus on feature extraction from raw signals without sample augmentation and often utilize single-scale convolutions to extract features from multi-band EEG signals. This approach fails to fully explore the correlations between different frequency bands. Therefore a sample augmentation and multi-scale decoding method was proposed based on the mainstream EEG-TCNet decoding technique. Methods First the motor imagery EEG signals were segmented to increase the number of samples in the dataset. The motor imagery EEG signals were downsampled at equal intervals into three different subsequences with each subsequence containing the same data characteristics as the original motor imagery EEG signal. Next EEGNet was used to extract features from each subsequence employing different scales of EEGNet for different subsequences to capture features from various frequency bands. Afterward a convolutional sliding method was applied to further segment each subsequence processed by EEGNet fully exploring the latent information of each subsequence. Subsequently each processed subsequence was fed into a temporal convolutional network for feature extraction and dimensionality reduction. Finally all processed subsequences were concatenated and averaged which were then input into a fully connected layer for recognition. Results The proposed improved network was validated on the public BCI Competition Dataset IV-2a showing an increase in decoding accuracy of 5. 19% and 7. 7% compared with EEG-TCNet and EEGNet respectively. Conclusion The improved network demonstrates better decoding performance in motor imagery EEG signal recognition tasks.