Face Recognition Based On PCA Algorithm And Face Pose Synthesis

Face recognition refers to the identification of a given face image from the known image database. Feature extraction is the core problem of face recognition. Principal Component Analysis (PCA) method for feature extraction is a successful linear analysis method with the advantages of high speed and high recognition rate for the frontal faces. But the disadvantage is that the robustness of recognition rate is easily affected by factors such as light illumination, face expression, and face posture. The paper focuses on the light illumination and face posture factors in face recognition.A pretreatment strategy that can deal with image gradation was developed to restrain upper sensitivity of PCA from the change of light illumination in the image. This paper proposed an efficient method for face recognition. The training and testing samples were selected with random sequences from the face database, and then processed with a power transformation and a Butterworth low-pass filter, and finally processed with the PCA algorithm. The experimental results based on the ORL database showed that the proposed algorithm achieved higher accuracy of recognition than the traditional ones if we selected the power transformation values properly.Although the PCA algorithm for the frontal face recognition has achieved rather good result, the recognition rate can drop sharply for some facial features will not display when the face posture changes. The research on face recognition algorithm dealing with the pose change is relatively scarce in literature. Therefore, the pose problem in face recognition is an urgent one to be solved.The synthesis of multi-pose face images is a recovery process of facial information. Based on the prior knowledge of limited profile face images and the pose change information, we incorporated the active appearance model to extract texture information for each posture and got aligned multi-pose face image. Then the face images were divided into the training and test sets, respectively. We proposed a generative model that created a one-to-many mapping from an idealized identity space to the observed data space. Then we used the expectation maximization algorithm to estimate the parameters from the training data, so that we could synthesize the faces under a new pose. The experimental results on the ORL databases showed the validity of our multi-pose synthesis method. The experimental results proved the effectiveness of our method.