A Study On Feather Extraction And Classification Algorithms In Face Recognition

The strong adaptability, high security and non-contact smartinteraction of face recognition make it of great potential in applicationssuch like public security, intelligent access control and criminal investiga-tion, and make it become one of the most popular topics in the field of pat-tern recognition and computer vision. Traditional face recognition systemgenerally contains four steps: Face detection, image preprocess, featureextraction, and classification with some classifier, among which featureextraction and classification are the cores.Around feather extraction and classification algorithms for face recog-nition, this paper has mainly made contributions on the following threeaspects:(1)Weobtainanewsupervisedfeatherextractionmethod, uncorrelateddiscriminant sparse preserving projection method, by adding discriminan-t information into sparse preserving projection. Through preserving thesparse within class and maximizing the distance between classes, the ob-tainedprojectioncannotonlyexpressdiscriminantinformationeffectively,but also preserve the local neighbour relationships. Furthermore, the un-correlated constraint can reduce the redundancy in feather vectors, whichhelps to get more feather information from less feather vectors.(2)In order to classify the image data, it is usually to convert them intovectors, which would result in the distortion of correlative information oftheelementsintheimage. Thispaperdesignsaclassifiercalledtwodimen-sional neural network with random weights (2D-NNRW). It can put matrixdata into the classifier directly, and can preserve the structure of images.Specifically, the proposed classifier employs the left and right projectingvectors to replace the usual high dimensional input weight in the hidden layer to keep the correlative information of the elements, and adopts theidea of neural network with random weights (NNRW) to learn all the pa-rameters quickly.(3)The problem of multi-pose face recognition with varying illumina-tion is studied based on the sparse representation. To make the sparse rep-resentationmoresuitablefor themulti-poseproblem, weproposeaweight-ed block sparse representation algorithm to emphasize the contributions ofthe similar poses to the representation of the test face, which improves therecognition rate of the system. Furthermore, in order to avoid adding moreillumination images into the training set, that is, solve the problem of rec-ognizing multi-pose face images with varying illumination under standardillumination condition, we propose the concept of illumination dictionary,and use it to sparsely represent the illumination variation in the test image.Moreover, we give a method for constructing the illumination dictionary.The corresponding experiments in each part verify that the proposedmethods all can improve the recognition rate to some extent.