Multi-task convolutional network for artwork analysis and personalized guidance

Traditional art education assessment is susceptible to subjective judgment and delayed feedback, while single visual models struggle to simultaneously achieve evaluation dimension recognition, quality scoring, and personalized guidance. This study constructs AI-CNN-Driven Innovation Model for University Fine Arts Education (AICDIM). The model employs Art-Element Aware Attention (AEAA), which extracts visual evidence corresponding to art evaluation rubrics through four branches, structure, chiaroscuro, color, and composition, while utilizing teacher-annotated regions to constrain attention localization. Teacher comments are encoded into textual semantic representations for visual-language alignment during the training phase. Multi-Task Learning (MTL) jointly accomplishes artwork category recognition, quality scoring, and artistic style identification. Q-learning with safety constraints selects instructional recommendations based on students’ historical performance. Experiments are conducted on WikiArt, OpenUArt-Student, and APDDv2 datasets. AICDIM achieves classification accuracy, style recognition F1-score, scoring MSE (Mean Squared Error), and defect detection mean average precision (mAP) of 93.8%, 0.914, 4.25, and 76.2% on OpenUArt-Student, respectively. On APDDv2, the overall scoring MSE and Pearson Linear Correlation Coefficient (PLCC) reach 0.62 and 0.86, respectively. The results demonstrate that rubric-aware attention, semantic constraints, and dynamic feedback strategies enhance the accuracy, interpretability, and pedagogical applicability of artwork evaluation.