Abstract
In the field of visual communication design, the aesthetic quality assessment of packaging images faces significant challenges due to the complexity and subjectivity of their layout composition. To enhance the objectivity and intelligence of such evaluations, this study proposes a packaging design aesthetic quality assessment method combining image composition features and graph neural networks (CGA-GNN). The method extracts visual structural information from images based on graph construction rules (e.g., symmetry, proximity, rule of thirds) and integrates a graph attention mechanism to improve compositional awareness in node feature aggregation. Experiments were conducted on the constructed dataset of 1,200 annotated packaging images. The results demonstrate that CGA-GNN significantly outperforms existing baseline models in both prediction accuracy and consistency. Specifically, the model achieves a Weighted Root Mean Squared Error (WRMSE) of 0.378 ± 0.018, which is significantly lower than that of GraphSAGE-GAT (0.397 ± 0.021, p < 0.05), GAT (0.425 ± 0.022, p < 0.01), and CNN (0.446 ± 0.023, p < 0.001). Regarding Spearman's rank correlation coefficient, CGA-GNN attains a score of 0.714 ± 0.017, Markedly higher than other comparative models, with a Maximum improvement of 0.073 (p < 0.001). Additionally, its Graph Structural Integrity Rate (GSIR) reaches 0.921 ± 0.016, representing an approximately 15% increase compared to CNN (0.802 ± 0.020). Ablation studies further reveal that the model achieves optimal performance when all three compositional rules are incorporated (WRMSE = 0.378, Spearman's ρ = 0.714, Kendall's W = 0.691), validating the complementary effect of multi-rule integration. The findings confirm the effectiveness of deep integration between composition rules and graph neural networks in assessing the aesthetic quality of packaging images, providing technical support for standardized design evaluation, personalized recommendation, and creative assistance.