Complex Emotion Estimation Using Analysis-by-Synthesis of Facial Expression Images

This research proposes an approach for recognizing facial expressions for complex emotion estimation through the utilization of generated facial images. The proposed approach consists of two main parts: facial image generation and facial expression recognition. In the first part, we introduce Conditioned Emotion Generative Adversarial Networks (cEmoGANs) to synthesize images that convey complex facial expressions. Unlike previous methods, our generative model maintains face identity information and expresses various types of complex emotions. This capability encourages the generator to have control over the image generation process, resulting in enhanced image quality, reduced distortion, and increased diversity of generated images. The second part involves the design of a multiple-label classification based on a convolutional neural network trained on the complex facial expression images generated from the first part, which is employed in the recognition of complex facial expressions for emotion estimation. Our model, using images generated by cEmoGANs, demonstrates a notable performance surpassing the capabilities of the previous models in comparative evaluations.