Non-uniform circular-structured loss inspired by psychology for image emotion recognition

In recent years, Image Emotion Recognition (IER) has attracted growing attention, primarily due to the rampant sharing of images on social networks. Unlike other computer vision tasks, IER poses significant challenges due to the subjectivity and complexity of emotion. Existing approaches primarily concentrate on IER model design, remaining a dearth of research on loss for IER. In this paper, we propose a Non-uniform Circular-Structured Loss (NCSLoss) inspired by psychology, which integrates the subjectivity and complexity of emotion into the training process, thereby enhancing model performance. Specifically, we construct a Non-uniform Proportional Emotion Circle (NPEC) based on the Russell’s theory, which possesses a more realistic representation of emotion distance relationships. Additionally, we propose Random Emotion Pointers (REP) to reduce the biases caused by sample personalization. Extensive experiments and comparisons have been conducted on two emotion datasets, demonstrating that the proposed method outperforms the state-of-the-art methods. Visualizations further prove the validity and interpretability of our method.