Optimization Of Cross-domain Image Classification Algorithm Based On Transfer Sparse Coding

With the continuous advancement of image acquisition technology,large-scale image data continues to emerge.It has become a hot research problem to use machine learning technology to automatically annotate,classify,and retrieve massive images.The traditional image classification often extracts the underlying features such as color,shape,and texture of the image.However,these low-level features have certain limitations and cannot be truly understood by human vision.Sparse coding can solve this problem well.Sparse coding is a typical classification method based on feature space.By matrix decomposition,the image is decomposed into a linear combination of a set of over-complete image dictionary and representation coefficients to obtain high-dimensional semantic features of sample data,which can overcome the problem of poor limitations of traditional algorithms.According to above principles,many algorithms has been proposed based on sparse coding.Among them,transfer sparse coding has achieved good results in image classification tasks.However,there are still some problems with transfer sparse coding,there are still some high-level semantic information is not considered,for example,geometries manifold structure information and label information between data,and the problem of nonlinear data obtain poor classification in original feature space.To solve above problems and improve the accuracy and robustness of cross-domain image classification,two improved algorithms are proposed based on transfer sparse coding.The main contributions of this thesis are listed as follows.(1)The image data samples are low-dimensional manifolds embed in high-dimensional Euclidean space.However,the traditional cross-domain image classification method ignores the spatial manifold structure of image data distribution.In addition,different labels in the source and target domains are often related in some semantic space,and all correspond to the same feature space,which are significance to the classification results.To solve the problems,it is necessary to integrate information of data coincidence degree and label correlation to measure the distribution difference between source domain and target domain.Based on this,a new regularization transfer sparse concept coding model that combines manifold learning and multi-label learning 'is proposed.By combining sparse coding and concept-based learning,which transfers cross-domain visual information and realize image classification.The experimental results show that the proposed method can obtain higher classification accuracy.(2)The proposed transfer sparse coding model is classified in the original space of the data,it is unable to effectively deal with the problem of highly nonlinear data.To solve the problem,inspired by the kernel method can effectively solve the problem of the linear separation of raw data,what's more,in order to reduce the excessive information loss in traditional image classification task and the reconstruction error of image features,we integrate the kernel method into the traditional linear algorithm to learn more robust sparse coding.Based on this,a kernel transfer sparse coding algorithm is proposed.The algorithm not only can efficiently deal with the data which not be separated by linearity,but also considers the local geometric structure of the data and minimizes the data distance between the domains,which reduces the reconstruction error.So it can accurately extract the common features of the source domain and the target domain,which is not only the characteristics of single domain.The experimental results show that the proposed method can obtain higher classification accuracy.