Apprentissage de métriques et méthodes à noyaux appliqués à la reconnaissance de personnes dans les images
Alexis Mignon · HAL (Le Centre pour la Communication Scientifique Directe) · 2012
Our work is devoted to person recognition in video images and focuses mainly on faces. We areinterested in the registration and recognition steps, assuming that the locations of faces in the images areknown.The registration step aims at compensating the location and pose variations of the faces, making themeasier to compare. We present a method to predict the location of key-points based on sparse regression.It predicts the offset between average and real positions of a key-point from the appearence of the imagearound the average positions.Our contributions to face recognition rely on the idea that two different representations of faces of the sameperson should be closer, with respect to a given distance measure, than those of two different persons. Wepropose a metric learning method that verifies these properties. Besides, the approach is general enoughto be able to learn a distance between different modalities.The models we use in our approaches are linear. To alleviate this limitation, they are extended to the nonlinearcase through the use of the kernel trick.A part of this thesis precisely deals with the properties of additive homogeneous kernels, well adapted forhistogram comparisons. We especially present some oringal theoretical results on the feature map of thepower mean kernel.