Face Recognition Research Based On Manifold Learning

This thesis discussed and researched the manifold learning algorithm in face recognition, and found the algorithms recognize rate have a huge decline while the face pictures has even a little pepper-and-salt noise .In order to solve this problem, use fractal encoding algorithm and local binary pattern algorithm improve locality preserving projection, get a algorithm which is more stable and closer to actual application.First ,because fractal code has a great simulation with gray image both in theory and distance compute ,propose use fractal code distance instead of Euclidean distance to construct neighborhood matrix ,the experiment results showed this algorithm not only has a 5% higher recognition rate while eigenvectors are reduced to 80 dimensions which the original algorithm has 100 dimensions but also prove there may be a manifold structure in human face fractal code space; Second ,based on first point and LBP`s great robustness, propose use LBP histogram replace Euclidean distance and finally got a feces recognition algorithm which has recognition rate about 50% higher than original LPP when face images include less than 10% salt-pepper noises, and the eigenvectors have reduce from 100 dimensions to 50 dimensions ,achieved target effect .