Face Recognition Based On RGB-D

Face recognition is a classic problem of computer vision;its essence is image classification.The recognition rate is the goal which the humanity pursues unceasingly.There are many problems in traditional face recognition,Such as change in light,head posture,foreign body occlusion.Although there are a lot of solutions for these problems,but there are always such conditions.So we must try our best to solve these problems.Aiming at these problems,we propose a face recognition method based on RGB-D.This method mainly solved two problems;both of them are changes in head posture and light.According to different head pose we divided images into subsets by use the random forest algorithm.To a certain extent,it solved the problem of recognition failure caused by posture change.Moreover,we use the infrared essence that not affected by the impact of light changes.the use of the depth image acquired by kinect for face recognition was also a good idea.There are our main works:First of all,obtaining images under different conditions,such as wide-range head posture changes,illuminations,facial expressions,sunglasses and hand occlusion from Kinect.And put RGB-D images of the same person under different conditions as an image set.Secondly,using k-means cluster algorithm to segment the image background.Then using the random forest algorithm to estimate the head pose and using our own algorithm for face detection and clipping.And then using the known head pose to divide face image into several subset images.For classification,a block-based covariance matrix was proposed to represent the subset image models on the Riemannian manifold to achieve the reduction.And then used the SVM model to learn each subset images separately and the final result was acquired by fusing the results of all subsets image.Finally,we designed an experiment by using the Biwi Kinect database to demonstrate the method proposed in this paper.The proposed algorithm has been evaluated on the dataset with more than 5000 RGB-D images in different conditions.The experimental results show that the proposed algorithm can achieve a recognition rate of up to 98.84%.