Research Of Domain Adaptation Methods Based On Deep Convolutional Neural Networks

With the rapid development of deep learning,deep convolutional neural networks have achieved great success in the field of computer vision.However,most deep convolutional neural networks need the support of massive label data,which brings about expensive and time-consuming data labeling problem.Therefore,it is very important to learn knowledge from large amounts of unlabeled data.In recent years,domain adaptation methods in transfer learning have successfully solved such problems.Domain adaptation builds classifiers with cross-domain recognition capability by transferring the knowledge of a dataset with rich labeled data(source domain)to a dataset with scarce or unavailable labels(target domain).This paper conducts related research on domain adaptation scenarios in different fields,and the main research results are as follows:(1)Aiming at the problem of partial domain adaptation,this paper proposes a partial domain adaptation algorithm which utilizes correlation diversity to transfer knowledge between different domains.The algorithm first constructs a subset selector by using the label distribution and feature distribution of the target domain images to avoid some source domain classes from participating in the domain adaptation process.For the remaining source domain images,the algorithm proposes a comprehensive weight calculation scheme to accurately quantify their transferability.This scheme combines the class-level and instance-level relevance of the source domain images from many aspects.Experiments on several benchmark datasets prove that the algorithm is superior to the latest existing deep domain adaptation methods and some domain adaptation methods.(2)Aiming at the problem of few-shot learning in domain adaptation,this paper introduces a universal few-shot learning of domain adaptation setting.Under this setting,the relationship between the source domain label set and the target domain label set is unknown.In order to solve this problem,this paper proposes a universal few-shot learning domain adaptation network.The network first designs corresponding domain adaptation methods for various possible situations in the setting,and secondly,the network is equipped with a novel class-level weighting mechanism to quantify the similarity between the source domain label set and the target domain label set,so as to gradually determine the relationship between the the source domain label set and the target domain label set and make the network choose the correct domain adaptation method for knowledge transfer.In this work,jigsaw puzzle are also used to expand the trainable images,and further solve the problem of fewer trainable images in few-shot learning.A series of experiments show the effectiveness and stability of the proposed network.