Image Experience Prediction for Historic Districts Using a CNN-Transformer Fusion Model

This study addresses fundamental challenges in historic district planning and design, specifically incorporating the emotional value of streetscape images into the design process. A deep learning-based sentiment analysis system was developed, utilising a convolution neural network (CNN) and transformer models to assess emotional tendencies and temporal states within images. The system employs a multi-view feature extraction framework, integrating VGG, ResNet CNNs, and the Swin Transformer model, resulting in a novel feature matrix. The attention mechanism and transfer learning strategy significantly enhance model accuracy in label recognition and classification. The primary contribution of this study is developing a novel multimodal fusion model, which markedly improves sentiment recognition accuracy and practical applicability. The application of this system to the Jiangnan Historic District underscores the enhanced appeal through the integration of emotional value. By identifying emotional tendencies in streetscape images, designers can make better-informed decisions that foster positive experiences. This research innovatively applies sentiment analysis to historic district design, highlighting the potential of the system to guide culturally sensitive and engaging urban planning. Our analysis of images from 12 Jiangnan historic districts demonstrated the efficiency of the system in aligning images with existing imaging libraries, offering valuable references and feedback. The results underscore the practical potential of deep learning in visual sentiment analysis and emphasize the significance of emotional value in enhancing experiences in historic districts. This study provides new insights and methodological support for planning and designing such areas.