This work proposes D-Inception, a method for emotion recognition based on the Inception neural network and Electroencephalographic (EEG) signal processing. D-Inception is divided into preprocessing, feature extraction, and classification layers. The preprocessing separates the EEG signal into the alpha, beta, and theta bands, and the feature extraction finds the power spectrum in the alpha band and spectrum entropy in the beta and theta bands. The classification layer is a simplified Inception that analyzes the features to find spectral, spatial, and local relations to categorize the EEG signals into happy, neutral, sad, and fearful emotions. The experiments were developed with DEAP and SEED datasets, and the results show that D-Inception achieves an average accuracy of 92% and generalizes the learning better than the methods recently proposed in the literature using the discrete model.