Low Resolution Face Recognition With Multi-pose Variations Based On Deep Learning

With the development of society, face recognition technology application is becoming more and more widely. In practical surveillance systems, the images captured often contain with low resolution and great pose variations. And the performance of the existing algorithms is not good at the low resolution images with pose variations. In order to solve the problems affected by pose variations and low resolution, a deep network structure that combines with deep belief network and extreme learning machines was proposed.In our method, the deep belief network is mainly used for exacting features, and extreme learning machines used for classification. The deep belief networks is constitute of multilayer restricted Boltzmann machine, learn by greedy learning layer by layer and fine tune by auto-encoder. The deep nonlinear network composed exhibit the excellent performance approach in terms of non-linear problems of complex function, effective learn the point-pairs relationship of the manifold assumption from high resolution and low resolution images, fully exploit the nonlinear contact of pose variations and the common features between high and low resolution images, then overcome the influences bring by the pose variations and low resolution to extract the essential characteristics of the input image. At the top of the deep belief network add extreme learning machines to classification. Finally, experiment on the UMIST database, ORL database and FERET database, which the influence of pose variations is relatively important. At the beginning of experiment, preprocessing the images in the database, and then through the preprocessed images training the network structure and classification, the final results show that the proposed method compared with other methods has advantages of high recognition rate and classification of short time.