Face Recognition Method Based On Estimation Of Edge Geometry

Effective expression of the image content is the basis for image recognition, so therepresentation and extraction of human face image feature is a basic problem for facerecognition. Edge is one of the basic features of the image, it is a part of an image featureis not continuous (or mutations) results, and it’s also an important foundation of imageanalysis technology, such as image segmentation, texture and shape feature extraction andso on. The edges of face image contain lots of facial organs and texture classificationinformation, it not only greatly reduces the quantity of human face image information tobe processed, but also protects the face contour structure not to be damaged.Generally, dealing with face edge in the area of face recognition is the traditionaledge detection algorithm, such as Wallis、Laplacian、Priwitt、Sobel and Rober algorithmand so on. These edge detection algorithm such as first-order derivatives and secondderivatives are all pixel-level. They are generally treating the gray remarkable changes inpoints of the neighborhood as a target to determine. Those edges are formed by the edgepoints, and they are the collection of pixels around the gray-scale intensity changes. In thisway it will break the two-dimensional line characteristic of the image itself, and therecognition result is not very good.In order to explore a better representation and estimation methods for the edgefeature of face to improve the recognition rate, this paper introduces Multi-scaleGeometric Analysis to the face recognition area and it is used to extract the edgegeometric features of face images. Then Quaternion Wavelet Transform is used to estimatethe edge geometry features to apply for the face recognition. And an edge geometryestimated face recognition algorithm is proposed which is based on Wedgelet andQuaternion Wavelet Transform. The method uses Wedgelet transform to extract the edgegeometric of facial images, and then uses the Quaternion Wavelet Transform to estimatethe edge geometric feature of faces and then gets the estimated matrix. Then Local BinaryPattern (LBP) is used to enhance the local texture about express ability of the estimatedmatrix. Finally use the nearest neighbor classifier to classify and obtain the recognitionrate. In this paper, we take a series of experiments on several face databases to verify thevalidity of the algorithm. For example, the face recognition algorithm is based onWedgelet and Quaternion Wavelet Transform whether it should or not and where to jointhe Local Binary Pattern to strengthen the texture features. The experimental results showthat joining the Local Binary Pattern after the estimated matrix to strengthen is the best.Make a contrast on the methods such as Wedgelet, Wavelet transform, Curvelet andContourlet to extract the edge geometric feature. The experiment results show that thewedgelet transform in the edge geometric extraction is superior to other transformations.Compare Gabor wavelet with Quaternion Wavelet on their estimating edge geometryability and the results show that Quaternion Wavelet Transform has lower computationalcomplexity and higher edge estimation ability. We take a classification experiment on thestandard 1996 FERET database, the proposed algorithm, on the Fb, Fc, DupI and DupIIsubset respectively achieved 86.44%, 76.29%, 55.96% and 48.29% recognition rate.Compared with the traditional edge detection algorithm, the proposed method is increasedby about 40 percentage points. The result shows the advantage of edge geometry estimatefeatures of the face recognition algorithm.This paper presented a new face recognition method based on edge geometry featureand it is superior to the traditional edge detection algorithms. The edge feature extractedby wedgelet transform is higher robust to the expressions, lighting, and aging changes offaces. The edge estimation of Quaternion Wavelet Transform is applying in facerecognition successfully and has a lower computational complexity. So the proposedmethod also has potential application value on the areas such as image fusion, human faceimage detection and image registration.