Dynamic Visual Design Algorithm Optimization and Innovation Based on Deep Learning

Background: In view of the weak static modeling ability and dynamic feature extraction efficiency of traditional dynamic visual design algorithms, this paper constructs an optimization framework based on deep learning. Methods: First, a Multi-scale Hierarchical Fusion Network (MHFN) is constructed. Through a multi-level feature extraction mechanism, the model can simultaneously capture key visual features at different semantic levels. Then the Self-Attention Module (SAM) is introduced to dynamically adjust the weight distribution using inter-frame similarity. Use generative adversarial network (GAN) to realize cross-domain visual generation. Results: MHFN scored 4.8 points and 4.7 points for perceived quality and style consistency (out of 5 points), respectively, and the running time was 35 milliseconds; the average accuracy rate of SAM in keyframe recognition reached 97.06%. Conclusion: Through the improvement of multidimensional algorithms, this research has significantly improved the ability of dynamic visual design in keyframe recognition and cross-domain generation.