Abstract
In recent years, steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) have attracted extensive attention owing to their high information transfer rate (ITR) and multi-target recognition capability. Deep learning (DL) methods have shown strong potential in SSVEP feature extraction and target frequency classification. However, their relatively high model complexity and dependence on large-scale calibration data increase the burden of data acquisition and may induce subject fatigue, thereby limiting user experience and practical application. To address these issues, this study proposes a Spatial and Channel Reconstruction-based Frequency-Time-Space Deep Convolutional Neural Network (SCR-FTSDNN) for efficient and accurate SSVEP target recognition. Specifically, the raw EEG signals are first transformed into the frequency domain using Fourier transform to extract preliminary frequency-related representations. Then, temporal convolution is employed to further capture dynamic temporal characteristics. In the network architecture, spatial and channel reconstruction convolution modules are introduced to enhance the extraction of discriminative features and improve feature representation capability. Moreover, dilated convolution is adopted to enlarge the receptive fields of frequency, spatial, and temporal features, suppress redundant information, and assign higher weights to key discriminative components. To evaluate the effectiveness of SCR-FTSDNN, systematic experiments are conducted on two public SSVEP datasets, namely Benchmark and BETA, and comparisons are made with several existing methods. The experimental results demonstrate that SCR-FTSDNN achieves superior recognition accuracy and robustness, particularly under limited calibration data, indicating its strong adaptability and potential for practical SSVEP-BCI applications. The reproducibility code is available at: https://github.com/Walek669/SSVEPclassification.