Multiple Information Fusion Face Recognition Research Based On Key Feature Points

With the development of face recognition technology for many years, themultimodal information fusion for face recognition has become today’smainstream direction. Using a single two-dimensional face image recognitionalgorithm is sensitive to environmental factors, such as illumination, expression,posture, and the recognition rate is low relatively. However, three-dimensionalface not only contains more space information but also has stronger robustness toenvironmental factors. Because of the complexity of three-dimensional face dataprocessing and calculation is higher, so face recognition algorithm that combineof two-dimensional face intensity image and3d face depth information imagenot only compute high efficiency, but also can get high recognition accuracy. Themain of this paper and innovation includes the following contents:This paper proposed a face recognition algorithms based on orthogonalsparse preserving projections of kernel for approximation problem about singletwo-dimensional face image using sparse preserving projection (SPP) to sparsereconstruction on the original sample. In order to get sparse representationcoefficients that contains more identification information by kernel method, itmapped samples to high-dimensional feature space. Then, reconstruct sparsecoefficient of kernel sparse representation, increase the similar non neighborsample weight, and reduce heterogeneous neighbor sample weight. Finally, thewhole orthogonal constraint transformation improve the ability of sparse retainsample. The algorithm experiments were carried out on the YALE_B and ORLface database, and the results verified the effectiveness and robustness of thealgorithm.In order to overcome the environment factors, this paper proposed apositioning algorithm of three-dimensional face data key feature points forgetting key feature points information of three-dimensional depth data. The algorithm uses median filter to remove isolated points for three-dimensional facemodel, it would make binary point cloud model, remove the background regioninformation locate the face region preliminary, and locate feature points. Theneliminate the interference point of the candidate points by calculating the pointdensity, and get the tip point. Finally according to the location of the nasal tippoint, we can get the key feature points on the horizontal and vertical direction.The algorithm can get rid of the interference information, and locate featurepoints accurately.According to the above method of getting key feature points matching andgenerating more complete information training samples. This paper proposes amultiple information fusion face recognition algorithm based on key featurepoints. It improves the sample clustering to reduce the nonlinear problem in theprocess of human face feature extraction by dividing the two-dimensional virtualimages into N subsets. Because the feature points are different about contributionto the classification ability under different pose and illumination conditions, weweighted processing on them. We can analysis and integrate key feature points oftwo-dimensional virtual image with LFA. Finally, using PCA algorithm reducesdimension to avoid redundancy of LFA algorithm, and get the final featuresubspace.