One New Constructing Multi-layer Pyramid Adjacency Graph Method for Locality Preserving Projections Based on HOG Features
Xiangchun Yu · Journal of Information and Computational Science · 2014
Locality Preserving Projections (LPP) is a representative of the classical dimensionality reduction algorithm based on constructing Adjacency Graph which plays a vital role in the performance of LPP. In the actual environment, face recognition will suffer the disturbance of occlusion and scale size of image shot. In order to reduce this bad effect, we propose One New constructing Adjacency Graph method named Multi-layer Pyramid Adjacency Graph Method (MPAG), which creates a five-layer pyramid for each face image and calculates neighborhoods on every layer. Besides, Histograms of Oriented Gradient (HOG) features possess a good invariance to local optical deformation and geometric transformations. The local shape is well characterized by the capturing edge or gradient structure. So, we apply the HOG features to face recognition, given our new method MPAG-LPP based on HOG features, which calculates neighborhoods on every layer based on HOG features. Several experiments are performed on well-known face databases, which show good robustness to change of face image size shot in practice and to occlusion by glasses or other object.