Research on the Restorative Perception of Jiangnan Garden Scenes Based on Computer Vision and Generative Artificial Intelligence

Jiangnan garden scenes represent the ideal human habitat in traditional Chinese culture, highlighting the harmonious coexistence of people and their environment while providing transindividual restorative experiences. Previous studies have not focused on the quantitative visual characteristics of Jiangnan garden scenes that enhance restorative perceptions. Additionally, research on the restorative perception of the Jiangnan garden scenes often faces challenges regarding its applicability in design practice, as well as issues concerning credibility and interpretability. To this end, this study uses social media images and self-captured images to create a dataset of Jiangnan garden scenes, obtains quantitative features of the visual elements within them by Semantic Segmentation through the Pyramid Scene Parsing Network. Subsequently, Deepseek-R1 is used to design descriptive prompts for Jiangnan garden scenes based on the identified features. And the Low-Rank Adaptation of Stable Diffusion XL is trained using the image dataset, which is then paired with the designed prompts to generate Jiangnan garden scenes that have restorative effects. Finally, the restorative perception of these scenes is evaluated through a perceptual scale. The results show that by using Computer Vision and Generative Artificial Intelligence, it is possible to effectively extract visual features and intelligently generate healing scenes. This method provides a reference for the paradigm shift in architectural design during the era of Artificial Intelligence. It also provides insights for improving the quality of the built environment and promoting emotional recovery in the public sphere.