Research On Multi-view Adaptive Semi-supervised Feature Selection Algorithm

In recent years,with the popularity of camera equipments and the rapid development of internet technologies,hundreds of millions of images are generated every day and spread rapidly in the internet.Massive images not only consume a lot of resources,but also are difficult to be searched and managed efficiently.Therefore,how to effectively manage these images becomes a hot issue in the image processing field and computer vision field.Images are often represented by high-dimensional feature vectors.High-dimensional data increases the requirements of the time and the space for processing them.In addition,“dimensional disasters” can occur sometimes and it makes data difficult to be processed.Therefore,feature selection techniques,especially semi-supervised feature selection(SSFS),have been extensively studied in the past decade.Among different SSFS methods,the semi-supervised feature selection method based on graph Laplacian is widely applied for its high accuracy.Although the graph-based semi-supervised feature selection algorithms achieve good performance in the fields of image processing and recognition,they still have their drawbacks.Firstly,the performance of these algorithms is heavily dependent on the quality of the Laplacian weight graphs.Secondly,when facing multi-view feature vector of images,these algorithms often directly concatenate the multi-view feature vector into a long vector,ignoring the complementary information contained in the multiple views.To overcome the above defects,this thesis proposes a Multi-view Adaptive Semisupervised Feature Selection algorithm(MASFS).The algorithm introduces self-paced learning into graph-based semi-supervised feature selection so that the graph can be adaptively updated according to the feedback information of current predicted label.Meanwhile,through multi-view learning,the complementary information contained in different views can be effectively utilized to improve the performance of feature selection.Then,an effective iterative method is proposed to solve the objective function.Finally,image annotation experiments are performed on NUS-WIDE dataset and MSRA-MM2.0dataset.Experimental results show that MASFS has better annotation performance compared with other graph-based semi-supervised feature selection algorithms.Figure 17;Table 7;Reference 59...