| The kernel method is essential for the support vector machine to solve nonlinear classification problems.It maps the data into a high-dimensional space,making the problem linearly separable.However,single-kernel SVMs may lack generalization ability and robustness for large datasets.This thesis investigates methods and techniques that combine multi-kernel learning and ensemble learning to improve the generalization and robustness of classification models.The main research contents of this thesis are as follows:First of all,the traditional multi-kernel learning algorithms consider only the combined structure of data kernel representation,ignoring common structure between samples,resulting in poor generalization and robustness.This thesis proposes a multi-kernel ensemble learning method for image classification using a shared parameter space.This method builds a multi-kernel ensemble learning model using ensemble loss and optimizes optimal parameters in multiple Reproducing Kernel Hilbert Spaces.It introduces shared and specific parameters to learn common and unique structures for each kernel,improving classification accuracy for large-scale data.Experimental results on UCI and image datasets demonstrate the effectiveness of this method over traditional multi-kernel learning.Second,building upon the previous model,this thesis proposes a multi-kernel ensemble learning method based on the solution space of multiple dual variables.By introducing the concept of multiple Lagrangian multipliers,each data sample has a corresponding Lagrangian multiplier enabling the optimal parameters to be solved in the solution space of multiple dual variables.The proposed method improves the accuracy and robustness of the multi-kernel ensemble model by transforming the constraint attributes and optimizing the parameters.Experimental results on image vision datasets such as LFW demonstrate its superiority over traditional multi-kernel learning methods in image classification tasks.Finally,based on the aforementioned research,this thesis designed and implemented a prototype system for image classification,which includes an image classification module and a user management module.Experimental results show that the system can effectively accomplish the task of image classification. |
