The Research Of Face Detection Based On Multiple Kernel Learning

As a novel classification method, though the sample size is large and themultidimensional data is irregular, multiple kernel learning can obtain goodclassification performance. Compared to other machine learning algorithms, in MKLframework, researchers could train different fuatures with choosing different kernelfunction, and achieved contribution (weighting) of each feature with the learningmethod. Though multiple kernel learning has been reflected many of its uniqueadvantage to solve classification problem, there was no uniform standard to effectivelychoose kernel function and its parameters, and determined the relationship of sparseMKL and non sparse MKL. Therefore, MKL was needed futher study.Face detection is the key of face recognition, has been become an indenpend andhot research issue. Face detection was defined as that determined whether a face in animage, and marked the location, size, number and position. In recent years, it has beenapplied to the content detection, authentication, security access control, security systemand so on. However, there were some challenges in face detection, such as multipleposition, light, expression and scale. And therefore, it was difficult to manual extractface factures.The research direction of this thesis was face detection based on MKL. It wasmainly addressed the issue of the limited kernel function selection of SVM and describeability of individual facture. This paper improved the existing multiple kernel learningmethods and multiple features combination. The main research contents of this paperwere described as follows.â‘ Described the theoretical basis of multiple kernel learning in detail, analyzed thebasic concept and nature of binary classification and kernel method, and given the solveapproach of multiple kernel support vector machine. To depth study of the main MKLalgorithms. To address a novel mix norm multiple kernel learning (MNMKL) andMNMKL gradient descrnt algorithm.â‘¡Outlined the face detection methods and current main algorithms, and combined the SIFT feature description. The first time, MNMKL was used to detection face.Then,analyzed the face detection algorithm based on MNMKL.â‘¢According to the influence of face features and the limited describe ability ofindividual feature, we put forward face detection algorithm based on multiple kernelmultiple feature(MKMF). This approach combined features which included LBP, RGBand SIFT, that is to say, described the image based on texture, color and local features.Then, trained these feature with linear kernel, Gassian kernel and2kernel, respectly.Next, decided the weight of single kernel with MKL.â‘£Vertifed the effectiveness of our proposed algorithms by experiments on theCaltech101standard image sets, MIT-CBC face databases and multiple faces imagefrom real life.