The assessment of art learning outcomes has long relied on teachers’ subjective judgment, facing challenges such as inconsistent evaluation criteria and difficulty in multi-dimensional quantitative analysis. To address these issues, this study proposes a framework for the automatic assessment of art learning outcomes based on symmetry analysis of multi-dimensional aesthetic features. The model quantifies the symmetry between student works and instructional exemplars across three aesthetic dimensions: color distribution features (HSV color space histograms and dominant color composition), compositional features (visual center distribution and structural symmetry), and art movement style features (multi-layer Gram matrices from VGG-19 with PCA dimensionality reduction). Using publicly available artwork datasets, this study constructed Temporal Evolution Pairs (early and late works by the same artist) and Stylistic Inheritance Pairs (works by different artists within the same movement) to validate the model’s effectiveness. The experimental results demonstrate that the proposed multi-dimensional feature fusion strategy achieves 87.6% accuracy in artist style evolution trajectory recognition and 82.3% accuracy in art movement style inheritance quantification, significantly outperforming baseline methods including SSIM (52.3%), VGG-fc features (68.9%), and single style loss (76.4%). Two in-depth case studies further validate the model’s quantitative capability: in analyzing Picasso’s stylistic evolution, the Mastery Index and the Creativity Divergence Index successfully captured the stylistic continuity of adjacent periods (Blue Period to Rose Period: the Mastery Index = 73.6) and the breakthrough innovation of cross-period transformations (Rose Period to Cubism: the Creativity Divergence Index = 82.7). t-SNE visualization of the feature space further revealed that deep style features can clearly distinguish different art movements and individual artists, with spatial distances between artists closely corresponding to stylistic affinities. This research provides new perspectives and tools for a computational framework with the potential for art education assessment practice. It is important to emphasize that the reported performance demonstrates the model’s ability to quantify stylistic relationships between artworks but does not yet demonstrate its validity for assessing student learning outcomes in real classroom settings. As noted, the current validation is based on art-historical consensus, and direct application to educational contexts will require further validation with actual student works and expert evaluation, which we plan to address in future work.
Ruan et al. (2026) studied this question.