Automatic seizure detection is of great significance to the diagnosis and treatment of patients with epilepsy. Due to the short duration of epileptic seizure period, the EEG signal distribution between the seizure period and the non-seizure period is imbalanced. To solve this problem, an automatic detection method of epilepsy based on the fusion of imbalanced classification and deep learning is proposed. Firstly, the Borderline-SMOTE algorithm is applied to one-third training set to prevent the boundaries between different classes from blurring. Then, a pyramidal one-dimensional convolutional neural network is designed, which is trained with the balanced processing data. Different from the common 2D convolutional neural network, the 1D convolutional neural network reduces the number of training parameters. The training rate is improved, and the overfitting is avoided effectively which is caused by the small number of training samples. By utilizing the 991 hours long scalp EEG database, the effectiveness of the seizure detection after balanced treatment is significantly improved. The sensitivity, specificity, positive predictive value, and negative predictive value reach 92. 35% , 99. 88% , 90. 68% , and 99. 91% , respectively. Meanwhile, the comparison with other seizure detection methods shows that the proposed method has better performance. It is suitable for satisfying requirements of clinical application.