Face Clustering Algorithm Based On Shape Features

The face recognition technology is a kind of biometric recognition technology. Ithas been widely used in many areas such as security and commercial because of itsconvenient, friendly use characteristics. The face recognition system involves largeamounts of data comparison operation. The recognition rate and speed have beenimproved greatly in a small database face recognition system to meet the basic real timerequirements. However under the condition of large-scale database, sequential searchstrategy makes the retrieval time longer, which is unsuitable for real-time application. Itis necessary to design a suitable structure for large-scale database to increase retrievalspeed.In this paper, clustering method and face recognition technology are combined toclassify face images according to facial shape features. Firstly, a face detection methodis used for face localization; and then facial shape features are extracted; thirdly, aclustering method is adopted to classify face database into several sub-sets according toface shape features; the further application of classification results is analyzed at last.The main contents are as follows:①A face detection algorithm based on AdaBoost is used to locate the position offace in image; the algorithm is suitable for real-time system for its high accuracy rateand speed.②An improved ASM method is used to extract facial shape feature points. In orderto better position the start shape, the initial face position is determined by a facedetector, moreover, a more better position is achieved by running two ASM searches inseries, using the result of the first search as the start shape for the second search. Atwo-dimensional profile model is proposed to improve the accuracy of landmark’slocalization; and the computation time is reduced by trimming the covariance matrix inthe process of calculating the Mahalanobis distance.③The ISODATA dynamic clustering algorithm based on Hausdorff distance isused for clustering of face images. A proposed algorithm improves the initialization ofcluster center value in clustering procedure.According to the experimental results, the face database is divided into sevenclasses reasonably and stably. In face recognition stage, the hierarchy structure canreduce the number of data matching, shorten the search time and increase the speed of recognition to meet the requirements of real-time application.