A Low-Energy Vision-Language Pattern Recognition Method for Design Style Analysis and Generative Evaluation in Art Design Education

Vollständiger Abstract

Worum geht es in dieser Arbeit?

To address the problems that style recognition of student works in art design education relies heavily on teachers experience, generative evaluation feedback tends to be generalized, and multimodal models usually involve high inference energy consumption, this paper proposes a low-energy vision-language pattern recognition method, namely LE-VLPR. The proposed method takes design images, work descriptions, task requirements, and teacher rubrics as inputs, constructs a design style prompt graph, and achieves alignment between image features and style semantics through vision-language contrastive learning. A lightweight adapter, dynamic visual Token selection, and 8-bit quantized inference are further introduced to reduce deployment cost. Meanwhile, rubric-constrained evaluation and triple consistency verification are incorporated to reduce feedback hallucinations. Experiments on ArtEdu-Style show that LE-VLPR achieves 90.28% Accuracy and 89.69% Macro-F1. For generative evaluation, the Pearson correlation reaches 0.846, the MAE is reduced to 4.37, and the hallucination rate decreases to 4.21%. The per-sample energy consumption is 1.83 J, and the EEP reaches 49.01. The experimental results indicate that the proposed method can achieve accurate style recognition and reliable instructional feedback under low-energy conditions.