Image Classification Based On Transfer Learning

Compared with traditional machine learning methods,transfer learning method has the ability to apply knowledge learned in source domain to the current task of target domain,and it does not require the distribution of training data is same as testing data.With the rapid popularization of electronic products,the number of images on the internet is growing exponentially.Some unknown or less known images have gradually appeared in human vision.Traditional machine learning methods may face on the problem of insufficient data supply and unreliability sample,when it is utilized for learning a classifier in target domain directly.The appearance of transfer learning offers a solution to solve these problems.And it has become one of the main concerns of many researchers.Aiming at the expression of transferred knowledge,this paper has discussed the transfer methods based on model parameter,feature and instance.The main contents as follows:(1).An improved LS-SVM algorithm based on model parameter transfer is proposed.The transfer based term is added to the original LS-SVM.It not only maximize the geometrical margin of LS-SVM,but also effectively employ the knowledge of the source domain to linear represent the model parameter of target domain.In order to use the multi-domain model parameters properly,we use the leave-one-out method to obtain the transferred weights of each source model parameter through various experiments.(2).We propose a supervised method which combines reconstruction weight and pairwise distance to obtain the manifold geometric structures in source feature space as transferred knowledge.We regard sophisticated feature and simple feature as source feature and target feature respectively.Because sophisticated feature has better performance in image description,while simple feature is opposite.The transferred knowledge could help targetfeature find a transformed space,where the mapped target feature still maintains the manifold geometric structures of source feature.Finally,The pairwise distance is controlled in certain range by label information to distinguish the different contribution from neighbors.(3).In this paper,there is a progressive relationship between weak classifier,weak-strong classifier and strong classifier.The paper utilizes this strategy to learn the strong classifier for target domain.In order to set correct weights for instance obtained from source domain,each source instance set combine with target instance set many times to learn some weak classifiers.We obtain a certain number of subsets of target instance sets by means of Bootstrapping algorithm.All the weak classifiers correspond to a source domain construct a weak-strong classifier.Finally,All the weak-strong classifiers construct a strong classifier.The way to update the weights for all instances is the same as Tr Ada Boost algorithm.The final strong classifier for target domain has obtained before the iteration end.