Emotion recognition based on electroencephalogram (EEG) is crucial for brain-computer interface (BCI) design. Previous studies typically acquire EEG signals from a large number of electrode channels to ensure a high accuracy, which may not meet the lightweight requirement for practical applications. In this paper, we propose a novel EEG super-resolution method to reconstruct high-resolution data from a limited number of electrode channels for emotion recognition. Specifically, the ResNet34 is mainly improved by removing maximum pooling layer from the neural network to retain the spectral features of EEG and modifying the fully connected layer of the residual network to accommodate the EEG data. An experimental-level batch normalization approach is used to further enhance the super-resolution performance. MSELoss is used as the loss function for model training. The experiments are conducted on the Shanghai Jiao Tong University Emotion EEG Dataset (SEED) with power spectral density (PSD) features as inputs. Multiple experiments are conducted to evaluate the super-resolution performance of our method. This study introduces for the first time a new validation experimental system that combines cross-individual validation trials to confirm the usefulness of produced super-resolution data. The results show that the super-resolution data generated based on 32-channel EEG data are similar to the corresponding 64-channel data, and emotion recognition can be realized based on EEG signals with fewer channels by using our method to reach similar recognition performance when using EEG signals with more channels. This paper is expected to provide evidences for the development of wearable EEG technologies in practical applications.