Epilepsy syndromes are typical childhood nervous system diseases that may include different types of epilepsy seizures commonly seen, but far more complex than seizures. Accurate classification of epilepsy syndromes is crucial for diagnosis and treatment. Scalp electroencephalogram (EEG) provides a favorable basis for clinical diagnosis of epilepsy syndrome. In this article, we present a comprehensive analysis on the correlation between time-/frequency-domain regional scalp EEG features and typical epilepsy syndromes, and propose a transfer network-based classification model for epilepsy syndromes. Results on 63 children suffered from four typical epilepsy syndromes and 19 children from the normal control groups (NCGs) show that: 1) the features of the frontal polar region and the frontal region are always very similar, and the parietal region and the occipital region have similar features for each syndrome; 2) skewness is the most significant feature and Lemp-Ziv complexity (LZC) has the least contribution to distinguishing childhood epilepsy syndromes/NCGs; and 3) different individuals of the same syndrome have similarities, while the EEG characteristics of different syndromes are significantly different. A ResNet50 model based on deep transfer feature learning is applied to perform epilepsy syndromes/NCGs classification. The results show that feature selection based on the feature significance testing and correlation analysis can well enhance the classification performance.