Face Recognition Base On Near-Infrared Images In Embedded System

Face recognition technology is an important part in biometric identification technology. The human face may not copy, capture conveniently, not need to be collaborated by user, which makes face recognition have a wide range of application prospects both in research area and in commercial market. And the embedded computing platform makes the recognition system applicable in mobile and portable area.This paper studies all the stages in the face recognition, discusses a series of process including imaging principle about face images, image preprocessing, common facial feature extraction method and mainstream face recognition algorithm. Theoretically speaking, near infrared image face recognition can resolve the difficult problem in light conditions change or masked face. Therefore, this system captures human face images by the near-infrared camera. This paper describes the characteristics of LBP features (local binary pattern) used in the field of face recognition, and improves general LBP operator to multi-scale. Experimental results show that multi-scale LBP operator LBP is better than the ordinary LBP. In the classifier design phase, the paper discusses the AdaBoost learning algorithm and how to get sample suitable for AdaBoost by conversed from multi-kind classification problem to two-kind classification. Meanwhile, it researches feature extraction and classifier match. The paper designed and implemented a fast and efficient face recognition system for embedded devices, using near-infrared light. Considering the memory limit and CPU speed in embedded devices, we need to choose the feature which can fully represent the characteristics of human face and the classification algorithm. Thence the system has completed used efficient LBP feature extraction and AdaBoost Classification Algorithm. For running in the embedded devices, the key algorithm code in the c language.For the face recognition system implemented in the article, this paper has done a lot of experiments for analyzing the performance and factors. The experiment results show some conclusion, there are as follow: image quality impact on the recognition rate; the more training samples of AdaBoost, the higher identification rate; Multi-scale LBP feature is superior than the single-scale; the runtime in test set can reach the real time recognition requirement in embedded system. While the face recognition system in this paper is designed for near-infrared image library, but also can be used for the visible light image library, which only need to re-train the AdaBoost classifier in visible light image library. Therefore, the face recognition system has good compatibility in different library.