Based On 3d Reconstruction Of Illuminationinvariant Frontal Face Image Synthesis

Face recognition is one of the hottest research topics in the field of computer vision and pattern recognition. For ideal pose and illumination of face recognition, existing methods have reached a very high recognition rate, when the pose or the illumination changes, regardless of identification method the recognition rate goes down, the visual effect will be worse as the same time. This thesis tries to study the posture and light processing technology based on a single face image, aiming at improving face recognition rate and the visual effect, the main work is as follows:Firstly, the 3D face data which are acquired is disorganized on the number of sampling points and different storage form. So the data normalization is a prerequisite for further study. The processing steps include face posture correction, area selection and sampling in different slice. In this paper, the normalized face contains 23676 vertexes, and storage in uniform vectors.Secondly, a method is proposed for frontal image synthesis from non-frontal. The method is based on the idea of statistical modeling, to reconstruct the missing face shape and texture. Firstly, 3D average model is applied to estimate the pose parameters of the test face image. Compressed sensing theory is used to filter prototype samples and then a more accurate model of deformation is built up. Secondly, the test face image is separately expressed by texture vector and shape vector. The deformation model theory is used to reconstruct front shape and texture. Finally synthesis texture was produced according to the original texture and reconstructed texture. The result shows that this method could synthesize natural frontal face image from non-frontal face image with effectiveness.Thirdly, a scale-based face image illumination normalization algorithm is proposed. The wavelet transform method is applied to decompose image into approximate low frequency part and high frequency part. The low frequency part of face image is mainly affected by illumination conditions and the face image texture information is mainly contained in the high frequency part. For approximate low frequencies, using light good half face based on the symmetry of the face returned to the overall quotient image. Solve self occlusion shading effectively, and reduce the amount of calculation. Besides enhance the high frequency part after noise reduction. Then the processed low frequency part and high frequency part are used to reconstruct face image according to inverse wavelet transform. The experimental results show that this method can achieve good visual effect.Finally, face recognition experiment is put forward to assess the performance which proposed in this thesis, the result shows that the method is effective in improving facial recognition.