Discriminative Optimization of Local Features for Face Recognition

The task of face recognition has been studied in the past few decades, which has led to introducing several methods. However, recently, the use of pictorial structures in object recognition has shown to be powerful. Face, also treated as a structured object, has parts (e.g. eyes, nose, and mouth) which are not dramatically deformed from one instance to another. This enables it to work fairly good on faces as well. Based on this fact, state-of-the -art face recognition methods detect facial features using pictorial structure approach and then apply some local feature descriptors in order to discriminatively train the recognition model. In this work, we propose an approach to optimize the local features at the detected facial points in terms of the patch size, constellation of cells, location, and normalization factors, using SVM Cross Validation Accuracy as the objective function and Gradient Ascent as searching method. We use HoG descriptor but the approach can be used for any other local descriptors. We test our approach on the task of face identification applying on a dataset of celebrity images which we downloaded automatically using Google search engine. We also report results of Face Verification task on the recently released standard dataset called Labelled Faces in the Wild (a.k.a. LFW).