In recent years, the development of Brain-Computer Interface (BCI) technology has garnered widespread attention and research interest. The core of BCI systems lies in their ability to accurately interpret brain intentions and translate these intentions into specific control commands to drive external devices or software systems. Among various BCI systems, Rapid Serial Visual Presentation (RSVP) stands out as an effective method for target detection and recognition, widely applied in scenarios such as visual attention tracking and cognitive load assessment. However, RSVP faces several challenges in Electroencephalogram (EEG) signal detection and recognition, including signal noise interference, significant individual differences, and insufficient model generality and robustness. To address these issues, this study proposes a novel end-to-end model architecture that references the Selective Kernel Networks (SKNet) to enhance EEGNet, aiming to improve the accuracy of EEG-based RSVP signal detection and the generalizability of the model. We adopt SKNet’s selective kernel mechanism to enhance the network structure of EEGNet, designing a specialized network for extracting spatio-temporal features from EEG signals. The selective kernel mechanism of SKNet dynamically adjusts the weights of different-sized convolutional kernels, enabling the network to flexibly capture multi-scale feature information. Through this approach, our model can more accurately identify Event-Related Potentials (ERPs) in RSVP tasks, along with other task-related EEG signal patterns. Experimental results on an RSVP-based EEG dataset demonstrate that our proposed method outperforms three other advanced analysis methods in terms of accuracy, Area Under the Curve (AUC), and F1-score evaluation metrics. These results not only validate the effectiveness of our proposed method but also provide new insights and methods for constructing BCI systems based on RSVP with broader applicability and higher recognition accuracy.