Research And Application Of Domain Adaptation Based On The Transfer Learning Method

Deep learning frameworks have made breakthrough progress in many computer vision tasks in recent years.However,some difficulties and challenges have also appeared in the actual application process,such as the larger requirements for the amount of training data,and the difference between the distribution of network training samples and actual examples.Therefore,the technology of transfer learning is introduced to solve the issue of data content during the network training process.However,due to the relatively poor interpretability of both deep network and transfer learning methods,there is currently no unified approach on how to use knowledge transfer methods in practical applications and how to maximize the role of transfer learning in real-world tasks.In summary,there is a lack of in-depth research.The lack of research on the mechanism of knowledge transfer in deep networks will lead to a lack of regular conclusions and methods in theoretical research.Meanwhile,a large amount of unnecessary resources will be consumed in the method trial and parameter adjustment process in practical applications.Aiming at the above problems,this thesis makes in-depth research on the domain adaptation mechanism based on transfer learning.This thesis tries to solve the problem of how to use transfer learning method when facing image recognition tasksin the real world,and proposes a series of dynamic optimization algorithms to reduce the resources consumed by the parameter adjustment process of knowledge transfer.At the same time,this thesis applied the regular conclusions obtained from the research and the proposed domain adaptation method to two real-world application scenarios,and both achieved good experimental results.The main work of this article is as follows:(1)Proposing the algorithms on selecting transfer layers and adjusting transfer parameters.A series of conclusions are summarized,including how to select the transfer layers,how to design the loss functions and how to achieve the balance between tranfer part and classification part.Meanwhile,two dynamic domain adaptive update algorithms are proposed to optimize the parameter update process of knowledge transfer,which greatly reduces the time and labor costs required for parameter adjustment.(2)Designing two reality dynamic transfer framework.The research on the domain adaptation mechanism and the proposed algorithm are applied to realistic scenarios.This thesis has selected two important application scenarios: cigarette laser code recognition and face recognition.This thesis demonstrated how to use the above-mentioned regular conclusions and related algorithms to guide the selection and application of knowledge transfer methods in the actual application.This thesis showed that the above conclusions and algorithms can not only reduce unnecessary resources invested in method selection,parameter adjustment,etc.,but also improve the final performance of vision-related tasks.