Block adjacency locality preserving projections based on hog for face recognition

Histograms of Oriented Gradient (HOG) descriptor is a local image descriptor, which shows good invariance to image geometry and optical deformation. The basic point of HOG is that the distribution of local intensity or edge directions can well characterize the local appearance and shape of images. We extract HOG features from face images as the features of faces. However, dimensionality of the feature is particularly high and contains redundant information. So, we propose a novel subspace feature learning method Block Adjacency Locality Preserving Projections (BALPP) based on HOG, which is a local description operator with a low dimensionality. This new approach has shown a certain degree of robustness to illumination and partial occlusion. We also present the results of experiment on well-known face databases, which perform well in accuracy and robustness. © 2015 ICIC International.