Research And Application Of Transfer Learning In Image Classification

With the rapidly development of the internet of things and images,the style of information representation is more and more diversified.Especially the image representation has the advantage of intuitive,easing to understand and so on.For example,every single news item of the news websites contains more than one image which makes the research and application of images become more and more important.If the problem of massive image annotation and classification can be solved,users can find the most valuable data conveniently and efficiently.However many images come from the distribution of different datasets and domains.Traditonal machine learning approaches usually assume that the training data and the test data follow the same distribution.If the traditional image annotation and classification algorithms are used to mine new data information,the performance must be degraded;if a single dataset is used to train which can not reflect the value of big data.In many real-world applications,it is difficult to meet these conditions.The performance of the model is greatly reduced even if the conditions are barely satisfied.With the rapidly development of computer information technology,how to exploit and use the effective information from massive data has been a hot issue.How to learn under the condition that the training data and the test data follow the different distributions is transfer learning.Transfer is a learning influence another which can occur not only in the field of knowledge and skills but also in motive,attitudes,behavior and interests.Transfer learning relaxes the constraint which the training data and the test data must be subject to the same distributions.Meanwhile transfer learning can mining essential structures and features in two different but related domains,and it can effectively share and transfer information between similar domains or tasks,so that the data can be transferred and reused in domain.Transfer learning has become a hot issue in data mining and machine learning.This thesis researches the problem of transfer learning,and puts forward two transfer learning algorithms based on a large number of research scholars.The contribution of this work is summarized as.follows:1.To the poor generalization ability of matching marginal probability distribution to reduce the domain difference,a feature joint probability distribution and instance-based transfer learning algorithm is proposed.Most of transfer learning methods are feature-based transfer learning and instance-based transfer learning.To improve the generalization ability of the model,this algorithm explores the new feature representation and the instance regularization term is added to choose relevant instances in source domain which further improve the transfer learning performance.The differences between the source domain and the target domain are obvious via feature-leaning and instance learning.The algorithm not only matches the marginal probability distribution but also matches the conditional probability distribution.The unified optimization objective function is proposed.The experiments on digital datasets and object recognition datasets show the effectiveness of the proposed algorithm.2.Most transfer learning approaches mainly focus on reducing the distribution different between source domain and target domain but ignore the potential semantic information,a transductive transfer learning approach based on manifold learning is proposed.The approach is a feature-based transfer learning approach which designs a new feature representation in low-dimensional space and maps the data from the original high-dimensional space to the new low-dimensional space.Specifically,the approach employs the manifold learning to discover the intrinsic semantic information hidden in the data space and simultaneously reduces the differences between the marginal distribution and the conditional distribution with different weights assigned.Experimental results on the common transfer learning datasets demonstrate the validity and effectiveness of the proposed approach.