AUGMENTED RANDOMIZATION INJECTION TRANSFER FRAMEWORK FOR FACE EXPRESSION RECOGNITION

In this paper we approach the theme of face recognition. Here difficulties arise due to the perceived subjectiveness of human observers, making the annotation process hard and costly. We propose a transfer based solution, in which the key element is the injection of a randomized perturbation within controlled amplitude for efficient regularization of the flow between two different domains, one with supervised data and one with unsupervised data. On the technical side, our method uses the self labeling paradigm and, as the images from the two cases, annotated and not annotated, may be drawn from biased distributions. To cope with the bias, a random perturbation is injected in the loss function while training. On the application side, to assess the efficiency of the proposed method we experiment two scenarios that have been rarely investigated before; these refer to the separability of anxiety -originated expressions in the wild and, respectively, to the study of face expression recognition in children.