Lightweight transfer learning for weathering pattern recognition in stone cultural heritage: validation at Yungang Grottoes
TL;DR
This study proposes a transfer learning framework using four lightweight pre-trained architectures (DenseNet-121, Inception-v3, MobileNet-v3-small, and ResNet-18), adapted through fine-tuning and iterative L1 pruning for efficient weathering pattern classification, providing an efficient framework for field monitoring in resource-constrained environments.
Abstract
Accurate identification of weathering patterns is crucial for monitoring and preserving stone cultural heritage. While computer vision techniques offer promising solutions for weathering pattern identification, practical challenges exist in field applications, particularly regarding model deployment efficiency and real-time processing capabilities. This study proposes a transfer learning framework using four lightweight pre-trained architectures (DenseNet-121, Inception-v3, MobileNet-v3-small, and ResNet-18), adapted through fine-tuning and iterative L1 pruning for efficient weathering pattern classification. Using the UNESCO World Heritage Site Yungang Grottoes as a case study, a dataset of 1000 images of five weathering patterns (powdery, plate, strip, sheet, and crack weathering) was constructed under varied field conditions. MobileNet-v3-small achieved the best classification performance with 94.15% precision and an F1-score of 94.02%, while maintaining a compact model size of only 5.95 MB. All models support real-time inference, with ResNet-18 reaching up to 5754.89 FPS. This study provides an efficient framework for field monitoring in resource-constrained environments.