Research on the recognition of cultural symbols in traditional residential building facades and their digital protection based on the LoRA-DenseNet model

Abstract

Digital preservation of traditional architectural heritage faces persistent challenges. To address these issues, this study proposes a lightweight framework that integrates Low-Rank Adaptation (LoRA) into DenseNet-121 for facade learning and cultural symbol recognition. Based on the concept of transfer learning, LoRA modules are embedded into key transition and classification layers of DenseNet, enabling fine-tuning of only 0.6% of parameters, thereby achieving efficient model training and feature expression. The study selects traditional residences in Jiangxi Province as the research subject. A small-sample image dataset with four distinct facade styles was constructed based on architectural characteristics and expert knowledge, upon which a series of empirical experiments were conducted. Experimental results demonstrate that the proposed model achieves a classification accuracy of 84% on the test set, significantly reduces computational resource consumption, effectively suppresses overfitting, and shows strong generalization ability and deployment adaptability. Grad-CAM visualization confirms that the model focuses on semantically meaningful components such as door buckets, gatehouses, and decorative carvings, aligning with architectural typology theory. Beyond technical performance, this approach reduces reliance on subjective annotation, improves interpretability, and enhances field applicability, thereby contributing to the scientific and scalable digital protection of traditional architecture.