EdgeSAM-CASD: Lightweight Mural Damage Segmentation via Convolutional Adapter

Jiansen Zhang, Zehua Wang, Wei Chen, Zishan Xu, Tingting Xu, Jueting Liu, Xin Li

Abstract: Digital preservation of cultural heritage demands efficient and precise mural damage segmentation. While the lightweight EdgeSAM(Edge Segment Anything Model) framework excels in edge device deployment, its capability to characterize multi-scale, irregular mural damage remains limited, and existing methods struggle to balance efficiency and accuracy. To address these challenges, we propose EdgeSAM-CASD, an enhanced EdgeSAM model incorporating a lightweight convolutional adapter (Conv-adapter). By integrating depthwise separable convolutions and a high-frequency retention mechanism, EdgeSAM-CASD strengthens feature extraction for damaged regions while reducing computational overhead. Our innovation lies in the introduction of a convolutional adapter integrated with a parameter-efficient fine-tuning strategy, enhancing the model’s feature extraction capability and task adaptability in complex mural damage scenarios. Experimental results demonstrate that EdgeSAM-CASD achieves a robust trade-off between efficiency and precision, offering a lightweight solution for intelligent mural conservation.

External IDs:dblp:conf/icic/ZhangWCXXLL25