Human Face Recognition Research Based On Lasso Extreme Learning Machine And LAB Color Space

For now, human face recognition is still a very difficult problem. Uneven illumination and instability is one of the problems in human face recognition systems, but it can be mostly solved by the mean light process in the CIE LAB color space. The RGB color space is relative with the equipment while the CIE LAB color space has no relationship with the equipment. When the image is transferred into the CIE LAB color space from the RGB color space, if making use of normalization of light from non-standard to standard in nature, it can prove the radio and affection of the recognition while the color information can not be lost. However, it can not widely used if the training speed of the neural network could not be proved because the method should use the neural network to recognize the human face. The classical neural network need long time to train the network which limits the application of the face recognition in the reality. Aimed to the problem of the long training time in classical neural network, this paper proposed a new algorithm for training the neural network named Lasso Extreme Learning Machine (Lasso-ELM) which is based the Lasso theory and the Extreme Learning Machine. The Extreme Learning Machine can decrease the training time by randomly assigning the parameters of the hidden layer in the network, but it need lots of hidden nodes and it has the overtraining problem. The Least Absolute Shrinkage and Select Operator (Lasso) can select and shrinkage the parameters. When the Lasso theory used in the Extreme Learning Machine, it can shrinkage and select the hidden nodes and it can also avoid the overtraining problem. This method proposed in this paper is based on wavelet decomposition analysis for extracting characteristic of human face which can reduce the complexity of computing and space of storage. The experiment in this paper is divided into two parts. The fist is the preprocedure which includes the color space transfer, even illumination and characteristic extraction by the db2wavelet. The second is respectively using the Lasso-ELM and the BP algorithm to training the neural network and simulation of recognizing the human face image. Finally, this paper will conclude that the radio of recognition can improved from82.5to88.7%, and the training time is reduced from27.6seconds to2.9seconds by using the Lasso-ELM, compared with using the BP network.