Research On Human Face Recognition Based On Local Features

Face recognition is a very important research direction in the field of computer vision and AI.And it plays an important role in various areas,such as public security and video monitoring system.Now,face recognition technology has developed rapidly,the researchers gradually do more study on image features,the local characteristics of image received widespread attention.Compare with the global feature of the image,the local feature performs better in grasping the details information and robustness.Local binary pattern(LBP)is an excellent local feature extraction method,which has good invariance of gray and rotation.It is an important method of face recognition technology.However,under complex scene,face image always suffers from large intra-class variations of illumination,appearance and pose.Therefore,the research of face recognition under complex scene is strongly needed.Detail works are as follows:Firstly,in general,face recognition technology include a series of steps from image acquisition,face detection,image preprocessing,feature extraction and classification.Each step plays a key role in face recognition.For face detection and image preprocessing,this paper mainly introduces the algorithm of face detection based on Haar-like and image pre-processing,which contains face slant correction,scale normalization,denoising,image pre-processing,and illumination pretreatment.These steps lay a foundation for the subsequent face recognition task.Secondly,aims at image local feature,this paper introduce the basic principle of the local binary pattern and its application in face recognition.Besides we took a lot of experiments on face image databases and analyze the advantages and disadvantages,which providing a theoretical foundation of face recognition method.Thirdly,to obtain the satisfied performance of face recognition under complex illumination,this paper proposes a method named face recognition based Multi-Direction Local Binary Pattern(MD-LBP).This method codes the face image by considering the relationship between each pixel of image and weighted pixels of its eight directions in its neighborhood.And then,divide the image after coding into several subblocks and take mean gray value of each subblock to compose the feature vector.The face image feature vector extracted by this method not only considers the relationship between each facial image pixels,but also describes the holistic spatial information of face image.Through lots of experiments,the experimental results show effectiveness of the proposed method.Finally,in view of the face recognition under the influence of expression and posture changes,the face recognition method was proposed based on Local Binary Pattern and the Convolutional Neural Network(CNN).This method using LBP to extracts the face image feature and gets the face image LBP feature maps.And then,take the face image LBP feature maps and original image together input CNN for training.The network structure parameters are adjusted by back propagation algorithm,and get the optimal face recognition network model.Through a large number of experiments on the standard face databases,experimental results show that the method in this paper can handle the problem of face recognition under the influence of the expression and posture changes,and achieve good recognition effect.