Convex Hull Algorithm And Nearest Subspace Analysis In Face Recognition Application

Along with the ceaseless development of social productivity, extracting feature andclassification become the basic activity. With the advent of the computer and theemergence of artificial intelligence, man will release from the repeating labor. Theseachievements are widely applied in many fields such as industry, agriculture, nationaldefense, scientific research, medical and health, meteorology, astronomy and etc. Afteranalysis of the achievements carefully, feature extraction becomes one of the core issuein partten recognition, in which the geometry and algebra feature is the importantmethods.We begin to research the convex hull and nearest subspace feature on2D faceimage. The main contribution is as follows,1. Proposed a fast convex hull algorithm for a large point set. After analysis of thepapers about convex hull, we found VAICH that algorithm exists some shortcomeswhich make it not remove more inner points and be suitable to all data sets. Based onthe analysis, the EM and EEM method is proposed solve the problem. And theexperiments on seven data sets prove that the proposed method is better than VAICH.As the convex hull can not extract the face feature under the illuminatio, occlusion andgesture efficiently, we only use the convex hull feature to classify the mass datasetcoarsely. At last, the experiment shows the results of face coase classification.2. Propsed a linear discriminat analysis method based on the global and nearestsubspace information. The idea is inspired by the theory of linear subspace and CV(Common Vector) method. We found that the methods, LRC (Linear RegressioClassification) and CV (Common Vector), only consider the relationship among its ownclass and the methods, LDA (Linear Discriminant Analysis), only consider therelationship among all classes. However, how to search a way to consider sigle classand global classes is our key point. And we present an objective function which makesthe distance among one class be near and the distance to left classes distant. Then, eachclass has its own orthogonality subspace. Otherwise, can the distance between samplesfrom different class subspace be compared? Next, we prove that the answer is yes. Theexperiment shows that the proposed method is better than the comparable ones. And, themethod is valuabe to further study on the problem.3. Proposed L1norm restriction on the coefficients of on sigle class to face recognition and sparse error. And, it can put the question of the face image underillumination, occlusion and noise into a processing frame. Inspired by the latest successof Sparse Representation-based Classification (SRC) in face classification, we adopt theclass-specified training samples to sparsely represent the probe samples and make theresidual meet the constraint of L1norm, to attain the aim of getting the relative correctresidual. To attain the aim, we proposed an objective function. In the process of solvingthe objective function, we use a trick to make the prolem be easy. At last, we attain thetwo sparse coefficients. In the paper, we give a suitable discriminant rule to lastclassification. Otherwise, the method can be also used in the other image basedrecognition.