Aims:Major depressive disorder (MDD) is a common psychiatric disorder, and objective measures that can support clinical assessment are increasingly being investigated. Electroencephalography (EEG) provides a non-invasive and relatively low-cost approach for examining brain activity and has shown potential for EEG-based MDD classification. This study investigated whether short resting-state EEG segments could distinguish individuals with MDD from healthy controls (HC) using two compact deep learning architectures: a Lightweight 1D Convolutional Neural Network (Light 1D-CNN) and a CNN-MiniTransformer. Materials and Methods: The study included 112 resting-state EEG recordings from 56 individuals with MDD and 56 HC. Ninety-two recordings were used for model development and internal evaluation, while 20 independent recordings were reserved for external validation. Nineteen-channel eyes-closed EEG recordings were standardized to 125 Hz and divided into non-overlapping 1-s segments. Channel-ablation analysis was performed to examine spatial EEG importance. Results: On the internal test set, the Light 1D-CNN achieved 96.43% accuracy, a 96.39% F1-score, and an AUROC of 0.9941, while the CNN-MiniTransformer achieved 95.26%, 95.31%, and 0.9906, respectively. On external validation, the Light 1D-CNN achieved 93.72% accuracy and an AUROC of 0.9785, whereas the CNN-MiniTransformer achieved 94.65% accuracy and an AUROC of 0.9872. P4 was the most influential channel in both global and MDD-specific analyses, with Fz and F4 also showing strong contributions. The parietal region showed the highest importance, followed by the frontal region. Conclusion: Both compact deep learning models showed strong performance for MDD–HC classification using short resting-state EEG segments and maintained high performance on independently held-out recordings. Similar channel-importance patterns across the two architectures also provide an interpretable basis for further investigation of spatial EEG characteristics associated with MDD.