Face Recognition Using LBP on an Image Transformation Based on Complex Network Degrees

Automated visual face recognition involves acquiring descriptive features from the image. Local Binary Patterns (LBP) is a powerful method to that end, capably characterizing local features. An crucial limitation of LBP, however, is that the feature vector's size becomes unmanageable when the method employed on even moderately large regions. In order to describe larger scale features, this work proposes a descriptor based on applying the LBP histogram applied to an image transformation based on node degree data derived from a complex network representation of the original image. The complex network generation heuristic and parameters are discussed. The complex network representation is shown to be able to condense larger scale image patterns into a local value that can be handled by LBP. LBP applied to this image transformation yields results that outperform LBP. We validate our proposed approach by applying our method to a face recognition task using three challenging databases. Results demonstrate that, for a large enough complex network generation radius, our method consistently outperforms LBP, while using a feature vector of the same size.