Depression is a common psychological disorder that significantly impairs individuals’ emotional health and daily functioning. Early identification and timely intervention are essential for effective treatment. This study proposes a deep learning-based classification method for depression detection using only electroencephalography (EEG) signals. Given the high temporal resolution of EEG, dynamic functional connectivity (DFC) is constructed to capture time-varying interactions between brain regions, enabling the joint modeling of temporal and spatial features. Resting-state EEG signals from the MODMA dataset are used, where subjects are divided into three categories—healthy, mild-to-moderate depression, and severe depression—according to Hamilton Depression Rating Scale (HAMD) scores. To extract discriminative features, power spectra from multiple frequency bands (delta, alpha, beta) are computed and projected into two-dimensional brain topography maps via azimuthal equidistant projection. These frequency-specific maps are fused into RGB images representing multi-dimensional EEG characteristics. The proposed model integrates a convolutional neural network (CNN) for spatial feature learning, a gated recurrent unit (GRU) network for temporal dependency modeling, and a graph convolutional network (GCN) to encode inter-channel topological relations. Experimental results show that the proposed method achieves a classification accuracy of 92.63%, demonstrating strong performance in distinguishing among different depression levels. This study highlights the feasibility of non-invasive, objective, and intelligent depression assessment based solely on EEG, and offers a promising direction for future mental health monitoring systems.