Research And Application Of The Combination Of Sparse Coding And Transfer Learning In Image Representation

With the coming of big data,everyone has to browse through a large amount of information in diverse medium every day.However,most of information is expressed by images.The image appears in our daily life frequently as it has the advantage of intuitive,easy to understand and the like,which makes the research of images plays an important role in our life.There are many applications referring to images including image annotation,image clustering,image retrieval and so forth.Therefore,how to solve the crucial procedure in these techniques of image processing is significant.Especially,we have to extract a "good" image representation from images.A great many methods of image representation are emerging including low level representation and high level representation.The high level representation of images has found such widespread use because it understands the semantic in images like human visual behavior even more.Like principal component analysis,sparse coding,non-negative matrix factorization,low rank representation.Whereas,these learning methods usually assume that the training dataset and test dataset has to follow the same distribution,which defects the generation performance of learned model in test dataset.Moreover,when extracting the high level semantic of images,we always estimate the reconstruction residual by the Gaussian distribution directly and ignore occlusions,corruptions,or useless information existed in natural images,which affects the robustness of the learned model.Transfer learning had come into being as a new technique of machine learning,which can effectively solve the problem of distribution divergence between training dataset and test dataset.As for image representation,a feature-based transfer learning method was proposed,which project the different dataset on the same feature space.The new feature representation not only describes semantic infonnation in images,but also reduces the distribution difference between the different dataset.The new characteristic of transfer learning simulates intelligent behavior of human as well as makes dataset from different domain related.In the era of big data,the transfer learning not only realizes knowledge sharing between different domains,but also annotates the new dataset by the old trained model,which saves the cost of manual image annotation greatly.Secondly,the assumption that the reconstruction residual follows the Gaussian distribution cannot deal with various outliers in daily life any more.Therefore,we try to fit the actual samples by adjusting the parameters in the probability density function of the residual according to the maximum likelihood in the statistics,which can improve the robustness of model.From the view of model learning,we introduce a noise matrix to capture the useless information in images automatically,which can weaken the interference of outliers to image representation.Inspired by those two points above,this dissertation proposes two distinct transfer learning algorithms based on the feature-based transfer learning.The contribution of this work is summarized as follows:From the view of the maximum likelihood in statistics,this dissertation proposes an image representation method based on transfer robust sparse learning.In this method,the weight matrix is used to fit residual distribution of the actual samples;the high level semantic content is obtained with sparse coding;the differences between the source images and target images are reduced by minimizing the maximum mean discrepancy;the geometrical properties in the image dataset are preserved by graph Laplacian matrix.There are two main innovations in this method.Firstly,the weight matrix reduces the influences to the coding learning and the dictionary learning by outliers.In the next place,the regularization parameter takes the place of the dictionary constraint of transfer sparse coding 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 some common transfer learning datasets show the effectiveness and robustness of the proposed method.From the view of model learning,as for the outliers in the actual images likewise,this dissertation proposes a transfer denoising sparse learning based on graph and joint distribution adaption for image representation.Different from the method above,This learning model captures the outliers automatically by a noise matrix,which reduces the labor to adjust weight matrix parameters.There are two main contributions in this method.One is that the noise matrix is employed to weaken the inference of outliers to the model learning.The other is that we further reduce the differences of conditional probability distribution between the different domains compared to original transfer sparse coding model.Experimental results on some common transfer learning datasets show the proposed method significantly improves the average accuracies in classification compared to other state-of-the-art transfer learning methods.