Research On The Method For Face Recognition Based On Extremely Randomized Trees

With the rapid development of science and technology, especially the familiarity of high-speed computer networks, people's lifestyle has undergone enormous changes and the degree of concerning about public safety are also increasing. After decades of development the measures for classification and recognition of the specific kinds of objects has reached a certain high level, but the application such as security, access control, transportation is endless. The ratio of object classification and recognition of existing methods is very accurate for the specific environment, under specific conditions, but also more demanding application asks for a higher challenge for existing methods.The main application object for the technology of object recognition is the human face. Based on its unique practical advantages and market positioning, face recognition technology is also concerned by industry and academic researchers.Objects identification of known types has become a research focus. Since the change of object's pose, illumination and scale, object recognition is especially difficult. When the object class is known, it becomes much more difficult to identify the object if the object is observed for the first time. Basing on the SIFT + ERT + SVM method of Eric, certain key aspects have been improved in this paper, and thus obtains a better recognition effect, with better promotion effect. There are three contributions of this paper to the image classification and recognition(comparing). Firstly, it changes the existing computing model of absolute Entropy to the relative entropy (Shannon entropy) mode ,increasing the stability of ERT(extremely randomized tree) and ERT's adaptability when generating the ERT, On the basis of the existing methods of using SIFT (scale invariant feature transfo- rm),adding ERT and SVM (support vector machine) in image classification and recogni- tion;Secondly,it carries a minimum entropy limit in the choice of randomized selecting splitting conditions on the basis of not losing the randomness of in this paper, obtaining better clustering results, thus making the method of this paper get better effect in classification and recognition , and also higher accuracy.