Face classification using quality weighted self organizing maps

Due to the weakness of the methods based on what we know or what we have, in recent years studies for identification have concentrated on the methods researching really who we are based on our biometrics. The fingerprint and the face have been the biometrics mostly used for this purpose. Previously, many face recognition systems were developed implementing deterministic or nondeterministic methods. The goal of the face recognition systems is, to identify the person on a still image or video using an existing database of faces. From this definition it is understood that face classification is a sub process located on the core of the face recognition process. For the cases in which the relations among observations are nonlinear, problem solving using the variations of Artificial Neural Networks (ANN) have been frequently utilized. The Self Organizing Map (SOM) which can successfully do the dimension reduction and build abstraction on relationships among observations, is a type of ANN used for pattern recognition and classification. In this study face classification is performed by a newly developed method using the SOM and according to the results obtained, the success of the proposed method is demonstrated by comparing with other methods. In the study, an application on Matlab was developed to obtain the results, the Yale B and the ORL face databases were used for tests. The application firstly performs the education of SOM using a specific percent of images then put the remaining faces to the test. Faces are processed after they are divided into user-defined and equal-sized blocks and the more important blocks picturing the face regions are more effective on classification result. The blocks at the same location on the face are classified among themselves and during determination of the best matching class, the class information of the neighboring blocks is taken into account and so that a safe, consistent classification handling cases like pose change, partial occlusions is aimed. Accuracy was evaluated according to the correctness of the class determined by the Euclidean distance between the best matching neuron position and positions of the class centers determined at training after the accommodating the face image to the best matching neuron. Classification successes of 91.67% for ORL face database and 86.10% for Yale B face database were obtained as a result of system run for 100 epochs , these values put forward the efficiency of the proposed method. In addition by the tests, the effects of dimensions of SOM, dimensions of the blocks, epoch count and neighbor effect coefficient on success were investigated and it is observed that there are no linear relationships between these factors and performance but there are optimum values of the factors for maximizing the performance.