Study On Key Issues In Automatic Face Recognition

Automatic face recognition is a hot study topic.Now under the illumination controlled and user cooperative conditions,automatic face recognition systems perform very well,but when it comes to the variant illumination,pose and expression(PIE) conditions,the performance of systems need to be improved.The recognition rate of face recognition systems in variant illumination condition degrade an order of magnitude compared with the invariant illumination condition.If other factors are considered at one time ,the recognition rate will degrade more seriously.The key problems of automatic face recognition field are as follows:(1).Face recognition with variant illumination,pose and expression; (2)A robust multi-angle of view face detection algorithm; (3)Extracting features of face images and the dimension reduction problems; (4)Analysis and modeling of face expression; (5)Designing of a robust automatic face recognition system.The dissertation researches on these key problems ,and several robust algorithms for face recognition with variant PIE have been proposed,to improve the recognition rate.1. The dissertation researches on Adaboost algorithm based on Haar-like features to detect faces fastly.The method combining YCbCr skin model detection and Adaboost is proposed, The skin detection method is used to validate the detection resultes obtained by Adaboost algorithm. It overcomes false detection problem of Adaboost . Experimental results show that nearly all non-face areas are removed,and the recognition rate is improved.2. An illumination processing method based on the framework of Retinex theory is proposed to reduce the effect of illumination variation. In the application of side-illumiation, classical Retinex algorithms enhanced the shadow edge falsely. In this paper, a novel Retinex algorithm based on adaptive smoothing with new conduction function is proposed .This conduction function uses both spatial gradient and local inhomogeneity to measure the severity extent of pixel variation. This function has no enhanced effect during smoothing, and it dose not lose the feature edge of face images. During iterative process, the maximum is chose between current iterative and last iterative ,which is used to be the constraint, so this adaptive smoothing method with proposed conduction function can be applied in Retinex theory to estimate illumination. Experimental results show that the proposed algorithm can overcome strong shadows efficiently without losing feature edges of face images.And the experimental results prove that the algorithm has illumination robustness ,it can improve recognition rate under any illumination condition effectively.3. The dissertation researches on methods of global feature analysis of face images and makes a investigation on the strategy of choosing different subspaces analysis with classifier. Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and Fast Independent Component Analysis (FastICA) are studied, and the advantages and disadvantages are compared. An optimal combination strategy is obtained by experiments.It is proved that a subspaces analysis method with the befitting classifier could obtain the optimal recognition performance.4. The dissertation researches on methods of local feature analysis of face images,the method"GLBP-PCA"is proposed .First,the method decomposes the normalized face image by convolving the face image with 5-scale and 8-orientation Gabor filters to extract its Gabor magnitude images.Then,the local binary patterns operates on each image to extract the local neighbor pattern.The histogram sequence extracted from all these region patterns is used to described the input face image and it is expressed as vector.In order to solve the problem of high-dimensional data,the PCA method is used to reduce dimension and recognition. Experimental results show that the GLBP-PCA method is better than using LBP operator only, and is not complex to compute. 5. SIFT algorithm is proposed to research on face recognition with variant PIE. Experimental results show that it could overcome the whole comparability of different faces and extract the local detail feature of face.The advantage of making use of SIFT algorithm is as follows: it can match and recognise face images of differ-size without normalizing face image complicatedly, and the algorithm dose not need training process,computing and experimentalizing are all simple. The experimental results, performed on three face databases with variant PIE, demonstrate the huge potential of SIFT algorithm in application to face recognition with variant PIE.6. An automatic face recognition system is designed.It can be used for identity recognition of faces in video frequency, and applied in many situations which are close set face recognition. The multi-face detection method combining Adaboost algorithm and YCbCr skin detection method ,and SIFT algorithm are core algorithms of system.A software system is done by VC,and it implement automatic face recognition of video frequency(including variant illumination situation) . Finally,a design plan of system frame based on DM6446 is proposed.