A multimodal deep learning-driven model for assessing drawing composition quality

Abstract

In recent years, digital art education, computerized evaluation systems, and aesthetic computing have prioritized drawing composition quality assessment. Traditional assessments have relied on subjective judgment or single-modal analysis, limiting scalability and uniformity. Existing methods use visual characteristics or rule-based models, which cannot incorporate multimodal inputs, including visual layout and verbal descriptions or rubrics, restricting score validity. This research proposes that multimodal Deep Learning-Driven Modal(MM-DLN) be used to create a reliable and intelligent automatic drawing composition assessment framework. Visual and textual data are integrated into a single model to increase scoring precision and interpretability. The technique uses MM-DLN, a Multimodal Transformer-Based Scoring Network. It uses a Vision Transformer (ViT) to extract spatial and compositional aspects from the drawing and a Large Language Model Meta AI (Version 3) (LLaMA3) to analyze textual input. A cross-modal attention mechanism fuses these modalities, and a regression-based scoring system generates the score. The model highlights attention regions, revealing its score logic. The proposed MM-DLN model reduced MAE by 23.9% and improved PCC by 10.2% compared to the best-performing baseline, demonstrating enhanced prediction accuracy and more substantial alignment with expert-driven aesthetic evaluation. Performance was stable with noisy or stylistically diversified inputs. In conclusion, MM-DLN uses multimodal deep learning to assess drawing composition quality automatically in an accurate, scalable, and interpretable method.