Depth-based face recognition by learning from 3D-LBP images

Abstract

In this paper, we propose a hybrid framework for face recognition from depth images, which is both effective and efficient. It consists of two main stages: First, the 3DLBP operator is applied to the raw depth data of the face, and used to build the corresponding descriptor images (DIs). However, such operator quantizes relative depth differences over/under ±7 to the same bin, so as to generate a fixed dimensional descriptor. To account for this behavior, we also propose a modification of the traditional operator that encodes depth differences using a sigmoid function. Then, a not-so-deep (shallow) convolutional neural network (SCNN) has been designed that learns from the DIs. This architecture showed two main advantages over the direct application of deep-CNN (DCNN) to depth images of the face: On the one hand, the DIs are capable of enriching the raw depth data, emphasizing relevant traits of the face, while reducing their acquisition noise. This resulted decisive in improving the learning capability of the network; On the other, the DIs capture low-level features of the face, thus playing the role for the SCNN as the first layers do in a DCNN architecture. In this way, the SCNN we have designed has much less layers and can be trained more easily and faster. Extensive experiments on low- and high-resolution depth face datasets confirmed us the above advantages, showing results that are comparable or superior to the state-of-the-art, using by far less training data, time, and memory occupancy of the network.