Deep Learning Landscape Evaluation System Integrating Poetic Emotion and Visual Features

This study integrates classical Chinese poetry’s textual and visual elements to quantitatively assess emotions in landscape designs and develop a corresponding evaluation system. A meticulously curated poetry dataset encompassing diverse emotional types and dynasties was selected. Utilizing Chinese-Contrastive Language-Image Pretraining (CN-CLIP) models, the visual features of the poetry were extracted, and a neural network model was trained to perform emotional analysis on landscape designs. The developed evaluation system combines a self-trained neural network with the Contrastive Language-Image Pretraining (CLIP) model, leveraging the emotions conveyed in poetry to assess landscape designs. An attention mechanism generates heatmaps that highlight the regions the model focuses on, enabling targeted design optimization. Experimental results demonstrate that the proposed “Emotion-Visual Features” model achieves an accuracy of 88.09% in emotion classification, significantly outperforming traditional methods such as Support Vector Machine (SVM) (64.51%) and Random Forest (RF) (81.47%), representing increases of approximately 36.5% and 8.1%, respectively. Additionally, the system’s evaluation module shows high consistency with expert manual evaluations, with system scoring improvements of 35.15% compared to 26.81% by experts, resulting in a minimal difference of 8.34%. These findings indicate that the deep learning-based evaluation system not only provides objective and quantitative emotional assessments but also effectively guides the optimization process in landscape design, validating the system’s reliability and effectiveness.