PERFORMANCE OF TRANSFER LEARNING FOR CRACK DETECTION OF MASONRY STRUCTURES USING YOLO11

Cross-domain generalization remains a key challenge in deep learning based crack detection of masonry structures, as most existing studies rely on a specific masonry type or a single data source, which limits the model adaptability to unseen scenarios. To address this issue, this study constructs a multi-source masonry surface crack dataset based on YOLO11 and systematically investigates the transfer learning combined with layer freezing strategies. Five training schemes are compared on a target domain to identify the configuration that best balances accuracy, generalization, and efficiency. These schemes are source-only, target-only, full transfer, backbone-frozen, and backbone-and-neck-frozen training. Results show that YOLO11 trained only on source-domain data achieves an AP50 of 45.0% on the target domain, indicating insufficient native generalization. Transfer learning raises AP50 to 94.1% and recall from 84.3% to 93.0%, while further freezing the backbone reduces the training time by approximately 46% at a comparable accuracy to the unfrozen transfer-learning model, with the highest stability across runs. These findings demonstrate that transfer learning combined with backbone freezing provides an effective and efficient strategy for improving the cross-domain performance of masonry crack detection models