Building a multimodal perception interaction system model for aesthetic education in higher education institutions using fusion algorithms

Based on an intelligent algorithm-driven multimodal perception system, this study constructs an interactive model for aesthetic education in higher education institutions, enabling collaborative recognition of visual, speech, and textual information. The model centers on a hierarchical attention fusion architecture, integrating ResNet-50, Bi-LSTM, and BERT feature encoding to achieve cross-modal semantic alignment and dynamic weight allocation. Experiments conducted on 18,000 sample sets demonstrate that the fusion model achieves a recognition accuracy of 95.4%, representing a 6.8% improvement over traditional Early Fusion models and an 18% reduction in response latency. It maintains 92.1% accuracy on cross-university datasets, showcasing strong generalization and stability.