Generation and Perception: A Computational Evaluation Method for Visual Quality and Emotional Impact in AI Artworks

Abstract

The rapid advancement of generative artificial intelligence has enabled significant breakthroughs in the visual realism and artistic expressiveness of AI-generated artworks. However, challenges persist in objectively evaluating their visual quality and emotional impact. Existing evaluation methods either focus on underlying image quality or rely on subjective user surveys, lacking a unified computational framework. This paper proposes a bimodal, multi-dimensional, and interpretable computational evaluation method. The Visual Quality and Emotion (VQ-Emo) framework is designed to jointly optimize visual quality scores and multidimensional emotional predictions. The framework comprises three core modules: a visual quality assessment module (a multi-branch convolutional neural network incorporating style perception, quantifying composition, color, texture, and lighting); an emotional impact computation module (a 20-dimensional multi-label emotion classifier based on the VAWE emotion model), and a visual-emotion association module (using attention mechanisms to identify emotion-driven visual regions). This model was trained and validated on the AGIQA-1K, ArtEmis, and self-constructed VAWE-Art datasets. Experimental results demonstrate that VQ-Emo outperforms existing methods in both visual quality assessment and emotion recognition, achieving an SRCC of 0.912 and a mAP of 0.678. Key contributions include proposing a multi-task learning framework for jointly optimizing visual quality and emotion, establishing the inaugural VAWE-Art dataset comprising 5,000 AI-generated images with 20-dimensional emotional annotations. This reveals quantifiable correlations between visual features and emotional responses across diverse artistic styles and provides computational foundations for emotion-controllable generative art systems.