Research On Kernel Based Image Recognition Algorithms

Image recognition is playing an increasingly important role in modern social life,while feature representation and feature modeling(classifier/regressor design)is the core task of image recognition.Due to the pattern complexity of the target and the background,the occlusions,and the influence of the shooting and lighting condition,the object often shows a varied appearance in the image.As a result,single feature and simple model often could not accurately depict the object in the image,and may leads to poor recognition rate.Multi-feature fusion is an effective way to solve this problem.This work have investigated the feature fusion methods based on the multi-kernel based metric and the multi-kernel support vector machine(SVM),as well as the post-fusion methods based on the integration of outputs from multiple kernel classifiers.The experimental system and results analysis are given offered for several image recognition problems.The main contributions of this paper are elaborated as follows:(1)A multiple kernel similarity metric(MKSM)is proposed to deal with the kinship verification problem using facial images.The MKSM is essentially the combination of multiple basic kernel similarity.Different from the previous methods for this problem,which are essentially using linear models,depending upon the choice of basic kernels,the MKSM can be either a linear model or a nonlinear one,or a combination of both.In this way,not only the model flexibility and capacity can be improved,but also the purpose of feature combination could be achieved.Moreover,a large margin(LM)criterion is proposed to optimize the proposed MKSM model,which can be converted into an optimization problem with L1 penalty upon the weight vector of kernel combination.This finally leads to sparse weight vector,thus achieves the feature selection in Kinship verification,and finally boost the efficiency at prediction stage.Experimental results on public datasets show that,using only small number of local features,the learned MKSM can match or outperform other the state-of-the-art methods.(2)A margin maximization method is proposed for multiple class kernel learn-ing.Two approaches are developed to learning the composite kernel,namely the k NN-based method(MCKM-k NN: multi-class kernel margin with knearest neighbors),and the sparse coding based method(MCKM-SR: multiclass kernel margin with sparse coding).Different from the previous margin motivated methods,in MCKM-k NN,only the k nearest neighbors in C are adopted to compute the distance between example x and class C,instead of taking all examples in C into use as the previous methods have done.Furthermore,different from MCKM-k NN,which use fixed number of neighbors and fixed neighbor weight 1/k for all examples,the MCKM-SR method determines the neighbor number and neighbor weights for each example by solving an sparse coding problem.Experimental results demonstrate that the MCKM-SR has significant improvement in comparison with the MCKMk NN,and both variant have show their superiority over the other benchmark methods.(3)A feature fusion method using SVM as transition functions is proposed to improve the robustness of hand detection under dynamic and cluttered environment.In this method,traditional and skin-Enhanced features are independently used to train SVM classifiers for each hand posture.Outputs of SVM for a single posture is collected and concatenated to train a logistic regression based classifier(LR)to perform this posture detection.And finally,based upon outputs of all SVMs,a softmax classifier is learned to further discriminate among all postures and hard backgrounds.Furthermore,a method using cascaded-softmax classifier is proposed to perform multiple class hand posture detection.It not only improves the detection accuracy and reduce the confusion rates between postures,but also have neat formulation which benefits the classifier training.Experimental results on a static hand posture dataset demonstrate the effectiveness of the proposed SVM+LR and the Cascaded-Softmax methods.