An Image Representation Method Based on Transfer Robust Sparse Coding
Peng ZHAO,Wei WANG,Hui-Ting LIU,Xia JI
DOI: https://doi.org/10.11897/SP.J.1016.2017.02421
2017-01-01
Chinese Journal of Computers
Abstract:Image feature representation method is one of the crucial problems in image process and image understanding.There have many important researches on low-level feature representation of images,such as HOG,SIFT,and so on.But there is still a big gap between the low-level feature representation and high-level semantic feature representation.So many machine learning methods are employed to learn the high-level feature representation of images for capturing the semantic information of images,such as Principle Component Analysis (PCA),Sparse Coding,non-negative matrix factorization,low-rank representation and so on.Traditional machine learning assumes that the labeled images and the unlabeled images are drawn from the same distribution and the residuals of image representation follow the Gaussian distribution.However,images update very quickly in the real world.The image generation environments are so diverse that the unlabeled images and the labeled images don't follow the same distribution.Moreover,there are often some outliers appearing in images,such as occlusion,corruption and so on.The feature representation residuals of the images can not be estimated with Gaussian distribution.So the data should be relabeled and the model should be retrained.Transfer learning methods allow the distribution of the labeled images (also named the training data) and the unlabeled images (also named the test data) to be different.The feature representation methods based on transfer learning aim to design a new “good” feature representation space,which can describe the semantic information of the labeled images and the unlabeled images better.In the new feature representation space,the statistic models (e.g.classification) learned from the training data (the labeled images) can be transferred to the test data (the unlabeled images).So the feature representation methods based on the transfer learning take full advantage of the labeled images and realize the knowledge transfer across image sets with different distributions.This paper proposes an image representation method based on transfer robust sparse coding (TRSC).In this method,the weight matrix is used to weaken the interference of outliers to classification.The high level semantic content of the image is obtained with sparse coding.The differences between the source images and the target images are reduced by minimizing the maximum mean discrepancy.The geometrical properties of the image database are preserved by graph Laplacian term.There are two main contributions in this paper.One is that the weight matrix is employed to generalize the residual distribution,which weakens the inference of outliers to the code and dictionary learning in the proposed image representation method based on Transfer Robust Sparse Coding (TRSC).The other is that the regularization parameter takes the place of the dictionary constraint of Transfer Sparse Coding (TSC) in the robust dictionary learning,which transfers the robust dictionary learning to the optimization problem and avoids the complexity of Lagrange solver.Experimental results on the common transfer learning databases show the proposed method significantly improves the average accuracies in classification compared to other 6 related representation methods,which proves the effectiveness and robustness of the proposed method.