Research On Cross Domain Image Classification Via Transfer Learning

With the number of images increasing,image classification becomes more and more important.Traditional image classification methods require large amount of labeled images.However,annotating images requires lots of labor and time,and is even unrealistic for some unusual or new categories.Few labeled images will lead to poor performance.Fortunately,we have many relevant labeled source domain images.How to use them to assist the target domain image classification is a hot problem in machine learning and computer vision field,which is also this paper's focus.This paper also focuses on how to overcome the large distribution shift between the source and target domains.This paper disposes cross domain image classification,focusing on target domain image classification tasks by transferring knowledge from source domain.First,we assume that there are only a few or no labeled RGB images in target domain,but there are a large number of labeled RGB images in source domain.In this case,we apply the source images to assist the target domain RGB image classification.Secondly,we assume that there are a few labeled RGB-D images in target domain,but there are a large number of labeled RGB images in source domain.In this case,we use the source RGB images to assist the target domain RGB-D image classification.Thirdly,in view of the advantages of depth images,we use the labeled source RGB-D images to assist the target RGB image classification.Finally,we assume that labels of target and source domains are disjoint,that is,zero-shot learning problem.We classify the target domain RGB images by the aid of source domain knowledge and the attributes shared by them.We study cross domain image classification via transfer learning in order to effectively transfer knowledge from source domain and improve the cross domain image classification performance.Specifically,the achievements and innovations are as follows:1.Given source domain contains a large number of labeled samples,target domain contains no labeled samples.This paper proposes a Projected Transfer Sparse Coding algorithm?PTSC?,to project source and target data into a shared low dimensional space and learn the sparse coding and a shared dictionary.But some source samples are still far away from target domain,even in the shared subspace.Source samples are assigned to different weights using L_2,1 norm constraint,according to their relations to target domain.The sparse representations are invariant to the distribution difference and the irrelevant samples.We learn the projection matrix,the sparse representations and the dictionary in a unified objective function.The results verify that the image representation learnt using PTSC improves target image classification performance and yields state-of-the-art results.2.Image data are high dimensional and nonlinear.Source domain contains a large number of labeled images,target domain contains a small amount of labeled images.Due to these characteristics,this paper proposes a Supervised Transfer Kernel Sparse Coding method?STKSC?,mapping source and target domain images to the kernel space for learning sparse coding and a shared dictionary.We also consider the distribution difference,the manifold structure and the correlation of the same class samples,to make the learnt sparse coding more discriminative.The results verify that STKSC can improve the target domain image classification accuracy.3.To cope with the problem of limited labeled RGB-D images,this paper proposes a method,Learning Coupled Classifiers with RGB images for RGB-D object recognition?LCCRRD?.We learn the coupled classifiers using source domain RGB images,the combined target domain RGB and depth images and target domain RGB images.The predicted results of the two target classifiers are made to be similar to make them more accurate.We also utilize the correlation between source and target RGB images making the shared relevant items boosted and the shared irrelevant items inhibited.The experimental results demonstrate that LCCRRD can achieve competing performance by utilizing relevant RGB images to help RGB-D image classification tasks.4.Depth images are more robust to illumination,complex back-ground,color changes and texture changes than RGB images.When the task is classifying RGB images,how to use the RGB-D images is important.To handle with this problem,the paper proposes a domain adaptation method by learning from RGB-D images to recognize RGB images,named DARDR.DARDR can maximize the correlations between RGB and depth images in source domain by low rank constrict and reduce the domain discrepancy across domains by the minimization of MMD.The merit of our method is that the correlation between source RGB and depth images and the discrepancy across domains can be incorporated with the classifiers of source and target data.Furthermore,a unified framework is presented to learn the classifier parameters.The experimental results verify that DARDR can utilize the relevant RGB-D images especially the depth information to help the RGB image classification task.5.Given that target domain has no training samples and has disjoint labels with source domain,which is zero-shot learning problem.Most existing zero-shot learning methods focus on how to project images into semantic space.The projection function learnt by source images and attributes has a shift for the prediction of target attributes.The paper proposes a method,zero-shot classification method by transferring knowledge from source domain and preserving target data structure?TKDS?.We learn the semantic correlation of seen and unseen classes and use it along with source classifier to learn target classifier.We simultaneously consider transferring knowledge from source domain and utilizing the manifold structure of target data to rectify domain shift problem.Extensive experiments show that the target classifier learnt by TKDS is more accurate,and achieves higher target domain image classification accuracy.