Machine Learning-Based Expression Generation Technology for Virtual Characters in Film and Television Art
Yi Zhang, Junting Qian
Abstract
Facial expression generation played an essential role in virtual avatars, television, film art, and human–computer interaction. The existing synthesis model faces difficulties due to poor generalization, unnatural distortion, identity inconsistency, and weak lip synchronization with audio. In addition, conventional approaches fail to balance expression fidelity, real-time synthesis, and identity preservation. The research difficulties are addressed by introducing the Machine Learning-based Expression Generation Model to improve the expression generation accuracy. The ML-EGM model uses spatial and temporal features to generate expression by integrating generative adversarial networks, transformer encoding, and lip synchronization. A generator synthesizes expressive face outputs in the generative framework, while the discriminator is trained to distinguish actual and created expressions. This adversarial process teaches the generator to mimic actual facial emotions. During this process, transformer encoding is applied to manage the temporal consistency using the self-attention mechanism to minimize the error and improve the similarity rate. Further, the expressions are integrated with the speech synthesis process to enhance the overall frame expression generation efficiency. The system uses the MEAD dataset to evaluate the system efficiency, in which the model attains 99.1% expression generation accuracy, which is a 5.7% improvement compared to the MFA methods. The ML-EGM model proposal sets a new standard in achieving realism in expression generation, consistency of emotion realism, and speed of synthesis; its possible use cases include virtual aides, AI avatars, deepfake image detection, robotics systems that convey expression, and emotion-sensitive human–computer systems.