摘要：Transfer learning can solve the problem of lowrecognition accuracy caused by dataset insufficiency. However,the improvement in performance for conventional transferlearning is limited. In this paper, we propose an iterativetransfer learning (ITL) framework to solve the problem. Usinga predetermined iteration strategy to perform ITL, the modelwith the best performance is selected to generate a new extendeddataset. The standard dataset and the extended dataset aremixed for training in the next round. This type of trainingprocess fully demonstrates the effects of transfer learning anddata amplification. The experimental results show that the ITLframework proposed in this paper improves the accuracy of theoptimal model from 91.63% to 97.70%. The ITL frameworkhas practical significance for improving model performance insmall datasets. It is suitable for the analysis of action sequenceson video streams with temporal characteristics and normativedefinitions.
关键词：Transfer learning; iterative model; ITL; data amplification; motion sequence analysis
