Research On Near Infrared Face Detection Based On Multiple-kernel Learning

Face detection is a complicated and challenging image classification problem.It is the key technology in face recognition system.The results of face detection directly restrict the effect of face recognition.It is actually a binary-classification problem to detect whether a face is present in a picture.SVM is the most commonly used classifier in image classification and target detection.However,the SVM has limitations in the large number of sample sizes and the uneven distribution of high dimensional space.The presentation of multiple kernel learning makes it possible for different features to be mapped with different kernel functions.MKL takes different kernels for different features,and trains the weights of each kernel according to different characteristics.It transforms the problem into the optimal combination of choosing kernel parameters and weight coefficients.We selected the optimal convex combination of kernel function to obtain optimal classification accuracy.Traditional face detection is greatly influenced by light in application.The technology has been used more and more because of the invariable characteristics of the near-infrared face on the change of light.In this paper,we proposed an improved MKL framework based on kernel parameters and kernel weight joint optimization.We mainly researched the problem of near-infrared face detection based on MKL framework.The main research contents are as follows:(1)Research and compare traditional cross-validation method and three common kernel similarity measure methods.The optimal one is selected to apply to proposed MKL framework in this paper.And,considering the problem of feature dimension,analyze Isomap algorithm,a fast Isomap algorithm is proposed to be applied to reduce the samples' feature dimension in proposed MKL framework.(2)This paper analyzed the problems in PSO in detail and an adaptive inertia weight and variable learning factor optimization method are proposed.Introducing the MKL theory systematically,and giving the common method of solving MKL.Study the current MKL algorithm,analyze the shortcomings and propose an improved method.In order to evaluate the performance of the proposed MKL framework,we used it to combine different features extracted from dataset.For the sake of comparison,the basis kernels constructed from the same features were combined by using the other state-of-the-art MKL algorithms.The results showed that this framework outperformed the other state-of-the-art MKL algorithms in terms of both classification accuracy and the computational time.(3)The proposed MKL framework is applied to a near-infrared face detection system based on Android.The test results show that this framework is effective and reasonable.