This paper presents a multi-modal emotion recognition framework integrating electroencephalography (EEG) and facial expressions. The architecture comprises four components: EEG processing, facial processing, dynamic weighting model, and a classification and constrained optimization. EEG processing involves spatial feature extraction via graph convolutional networks (GCN) to model electrode connectivity, followed by temporal modeling with bidirectional long short-term memory networks (Bi-LSTM) applied to 500 ms sliding windows of event-related potentials (ERP). Facial processing employs spatial attention mechanisms to focus on action unit (AU) regions critical for emotion expression, followed by dilated temporal convolutional networks (Dilated TCN) for sequential movement analysis. A dual-channel attention network dynamically computes modality-specific weights, fuse EEG and facial features through weighted concatenation. To minimize redundancy, mutual information regularization is incorporated into the loss function, which combines cross-entropy and mutual information constraints. Experiments on benchmark datasets demonstrate superior performance over state-of-the-art methods, validating the framework’s effectiveness in real-world applications.