Multi-Scale Facial Expression Generation Algorithm Based on Spatial Attention Mechanism

Abstract

With the maturation and application of human-computer interaction and virtual reality technologies, facial expression generation has received widespread attention. However, current technologies often fail to fully capture the fine details of the face, neglecting the differences between various facial parts in an expression, which lowers the quality of the generated expressions. To address this, this paper proposes a facial expression generation algorithm based on spatial attention. The algorithm combines spatial attention mechanisms with Generative Adversarial Networks (GANs), allowing the adversarial network to better focus on important facial details in the image. Additionally, the algorithm effectively addresses the phenomenon of feature differences across multiple scales by introducing convolution layers with varying receptive fields. This forces the neural network to learn facial features at different scales and the differences between them, thereby improving the quality of expression generation. Experimental results demonstrate that the proposed algorithm outperforms current mainstream methods in terms of expression generation quality, enhancing both the efficiency and quality of expression generation and providing new insights for facial expression research.