Multiple Kernel Mapping Based Image Feature Extraction Algorithms

Image feature extraction is one of the major issue in machine learning andmodel recognition tasks, including digital image classification, recognition,detection. While these modern tasks cannot be satisfied by classic linear methods,kernel learning based feature extractions, which are capable of implicitly mappingnonlinear feature spaces, are more and more employed and developed. However,single kernel based methods, constrainted by heterogeneous informations containedin data sourcesand limitation of specified to one certain type of features, often leadto unsatisfied effects when handling complex and high-precision required tasks.Therefore, multiple kernel mapping based image feature extraction gains increasingattentions from scholars. The main content of this paper is as followed.In technical system aspect, this paper focuses on the study of multiple kernelmapping based image feature extraction framework, and the implementation of thisframework to develop a image classification agorithm, and also presents acomprehensive discussion about the major factors that influence the agorithm effect.This paper is developed by providing a detailed introduction of graph embedding,and then combine it with multiple kernel mapping. With the excellent inductiveability of graph embedding, this paper expands two feature extraction agorithms.In method study, this paper develops comprehensive study on several mainfactors that influencesthe agorithm effect, i.e., function parameters, weights and thecombinations of basic kernels. Then several optimization methods, each respects tocertain factors, are provided and also combined, to maxmize the optimization effect.In experimental verification, to provide a sufficient comparison between theagorithm learned in this paper and classic ones, this paper employs the formermentioned framework to develop both multiple kernel and single kernel imageclassification agorithms, and implements them for face recognition, objectclassification and image segmentation on several databases. A comprehensiveanalysis is presented after comparing results, which indicates that the proposedagorithm is more precise than single kernel based agorithms. At last, a furtheranalysis focuses on factors that impact the computational efficiency, and severalcorresponding improvement methods are proposed.