Research On Feature Learning On Cross-view Data And Its Applications To Classification

Recently,data in various fields has exploded and an object can also be presented as different forms in various applications along with the rapid development of science and technology.However,the data of the same object in different domains plays a stimulative role in interpreting this object,which is more conducive to classify and recognize.However,in the face of cross-domian data,not only the same domain data has similar distributions,but also cross-domian data has semantic gap due to differences between domains,which leads that data of different classes in the same domain is more similar than data in different domains.Therefore,specific to the above problems,for the application of cross-domain data classification,this thesis researches the feature learning method of cross-domain data,and learns more discriminative inter-domain consistency features by designing some constraints for cross-domain data association mining.Moreover,three cross-domain data feature learning algorithms are proposed.The main research contents are as follows:(1)A cross-view feature learning method based on semantic association information is proposed to learn cross-view low-dimensional invariant features.In order to unlock the potential class structure and view structure of cross-view data,a self-representation model based on dual low-rank constraints is explored,which can separate two manifold structures in the learned subspace.In addition,two discriminant graph constraints are used to guide the affinity of the data in the above two manifolds respectively so as to reduce the influence of the view differences on the discrimination ability.Finally,an associated semantic common constraint is designed to learn view-shared and view-specific information from cross-view data in the semantic space so as to make the learned feature space capture complementary information between views and the view-invariant characteristics like semantic space.Furthermore,experiments on several public cross-view datasets prove that the proposed method is superior to some existing excellent feature learning algorithms.(2)A cross-view subspace learning method based on independent structure discrimination constraints is proposed.This method uses dual low-rank deconstruction constraints to decompose the class structure layer and view structure layer of cross-view data from the data space.In addition,for the separate class-independent structure and view-independent structure,the intra-view global discriminant constraint and the intra-view class global discriminant constraint are used to make the cross-view data align globally in each independent structure.Finally,several advanced subspace learning methods are selected to compare with the proposed method.The experimental results on four cross-view datasets and their data with different degrees of noise show that the proposed method is more effective and has stronger robustness.(3)A discrete cross-modal hash learning algorithm based on reverse nonlinear discriminant constraints is proposed to learn the public binary hash code of cross-modal data.In order to the task of cross-modal data retrieval,the proposed model uses a nonlinear embedded learning algorithm to compile the two types of modal data into a discrete common hash code.In addition,through a reverse regression constraint,the semantic information is regressed to the generated hash code,which not only enables the learned binary hash code to be used as the representative feature of discriminative classification,but also avoids the relax or the difficult of gradual compilation during the solution.The experimental MAP scores on three cross-modal datasets show that this algorithm is superior and more efficient than other advanced cross-modal retrieval algorithms.