A Novel End-to-End Neural Network for SSVEP-Based Brain-Computer Interfaces

Neural networks show significant promise in addressing recognition tasks conducted by brain-computer interfaces (BCIs) within the steady-state visual evoked potential (SSVEP) paradigm, but they still encounter challenges in tasks involving a larger number of targets, particularly in end-to-end manners. Therefore, this paper proposes an end-to-end time-frequency domain joint network, namely TF-SSVEPNet, for recognizing multi-channel SSVEP signals with different data lengths in the dataset with larger targets. TF-SSVEPNet consists of three modules: time-domain extraction, time-frequency conversion, and frequency-domain extraction. Experiments on the benchmark open-source dataset have verified that TF-SSVEPNet exhibits state-of-the-art performance in both user-dependent and user-independent training tasks in shorter gaze times. With a gaze time of no more than one second, TF-SSVEPNet achieves accuracies of 86.60% and 74.30% in two types of tasks, and information transmission rates reach 170.14 bpm and 132.43 bpm, respectively. This network structure improves the accuracy and efficiency of SSVEP recognition, which is beneficial for the practical promotion of BCI systems.