Ager_origial_dataset

Visual emotion analysis (VEA) aims to examine the visual stimuli that trigger emotional responses.Psychological studies suggest that such stimuli are not isolated but are often evoked through the complex interrelations among multiple visual elements within an image.Psychological studies indicate that emotions are often triggered by complex interactions among various emotional attributes within an image. However, existing methods primarily extract local representations rather than explicit attribute representations, and they lack reasoning mechanisms based on attribute correlations to explain emotional states.To address these issues, we propose Attribute-Guided Emotion Reasoning (AGER), a method designed to interpret emotional triggers through fine-grained attribute reasoning. AGER incorporates a hierarchical text-guided module and a multi-level expert network to extract fine-grained and high-quality attribute representations. Furthermore, we design an Aware Reasoning Module that analyzes complex relationships among attributes and their associations with emotions to explain emotional elicitation.Extensive experiments demonstrate the effectiveness of our method in emotion perception, and visualizations of fine-grained emotional attributes further verify the interpretability of our approach.