The rapid development of AI image generation technology has created an urgent need for systematic aesthetic evaluation of generated visual content. Existing computer vision assessment methods are often limited to technical image parameters and insufficiently consider multidimensional artistic judgment, texture details, and style consistency. This study develops a multidimensional aesthetic evaluation framework that decomposes image assessment into five core dimensions: composition balance, color harmony, theme expression, detail completeness, and style consistency. An attention-enhanced convolutional neural network is combined with transfer learning and multi-task learning to improve cross-domain generalization under limited labeled data conditions. To address the subjectivity of aesthetic preference and cultural differences, the optimization strategy introduces user-feedback-based online learning and interpretability analysis, enabling evaluation results to combine objective quantitative foundations with adaptability to specific application scenarios. Experimental results show that the proposed multidimensional fusion model achieves a Pearson correlation coefficient of 0.847 with human expert scores, outperforming single-dimension evaluation methods. The model provides technical support for AI-generated image quality control and offers methodological references for visual feature extraction, texture analysis, and electromagnetic imaging evaluation in engineering applications.
Yang et al. (Thu,) studied this question.