A Refinement Method for Optimizing Emotion Recognition Vector Clustering

Facial emotion recognition (FER) identifies human emotions through facial muscle movements, and research utilizing convolutional neural network (CNN) for FER has been actively conducted in recent years. However, CNN-based FER models are highly dependent on the quality of training data, which can be degraded by label inconsistency and inaccurate annotations. To mitigate these issues, a clustering-based approach is required to minimize human subjectivity and objectively interpret the intrinsic structure of the data. This study proposes an automated refinement pipeline that transforms high-dimensional emotion recognition vectors into a clustering-optimized representation. Experimental results using the FER-2013 dataset show that the refined data achieved a maximum clustering stability score of 0.92, while the unrefined data remained near zero, demonstrating the effectiveness of the proposed approach.