This study proposes a novel time-frequency dual-stream Electroencephalography Network (EEGNet) deep learning model for the automatic detection of epileptic seizures in Electroencephalography (EEG) signals. The model innovatively integrates temporal and frequency features and dynamically adjusts the contribution of each feature stream through an adaptive weighting mechanism, thereby effectively improving the accuracy and robustness of seizure detection. In addition, this study addresses the class imbalance problem in epilepsy detection by introducing clinically realistic data augmentation and focal loss strategies. Experimental results demonstrate that the proposed method achieves an accuracy of 98.27% and an F1 score of 93.77%, significantly outperforming existing methods. Ablation experiments show that both temporal and frequency streams contribute significantly to model performance, with the removal of the frequency stream resulting in a 6.46% decrease in F1 score. Confusion matrix analysis indicates that the model performs exceptionally well in distinguishing between seizure and non-seizure states, providing a new technical solution for automated epileptic seizure detection.