Few−shot learning of frame structure damage detection based on Meta−Learning and DCCMN model

Abstract

Abstract In response to the poor detection performance of frame structure under limited data conditions, this paper proposes a novel approach. This approach is based on the concepts of dynamic convolution and models such as ResNet and ShuffleNet. It introduces a Cross − Mix module and builds upon it to formulate the DCCMN (Dynamic Convolution Cross − Mix Network) model. Meta − Learning and the DCCMN model are applied to detect frame structure damage under few − shot scenarios (Meta + DCCMN). Experiments are conducted on the floor frame structure of Columbia University to validate the effectiveness of this approach. The proposed approach is subjected to N − way K − shot experiments and compared under the same conditions with SVM, ResNet − 18, DCCMN, and Meta + ResNet − 18 models. Experimental results demonstrate that, in the case of few − shot learning, the accuracy of this approach can reach 100% in 2 − way 5 − shot and 9 − way 10 − shot scenarios. Furthermore, the proposed damage detection approach outperforms other models, effectively addressing frame structure detection challenges under limited data conditions.