Efficient Hybrid Descriptor for Face Verification in the Wild Using the Deep Learning Approach

Resumo

In this work, we propose a novel model-based on a new Deep Hybrid Descriptor learning called DeepGLBSIF (Gabor Local Binarized Statistical Image Feature) for effective extraction and over-complete features in multilayer hierarchy. The typology of our methodology is the same as that of Convolutional Neural Network (CNN) which is one of the intensively-applied deep learning architectures. This field was developed due to: (i) end-to-end learning of the process utilizing a convolutional neural network (CNN), and (ii) the presence of very wide training databases. Our method allows improving the use of the interactions between global and local features for the model, which allowed providing effective and discriminating representations. In our study, the trainable kernels were substituted by our hybrid descriptor GLBSIF. Thus, the developed DeepGLBSIF architecture was efficiently and simply constructed and learned for Face Verification in the Wild. Finally, the classification process was carried out by applying distance measure Cosine and Support Vector Machine (SVM). Our experiments were performed on three large, real-world face datasets: LFW, PubFig and VGGface2. Experimental results demonstrate that our DeepGLBSIF approach provided competitive performance, compared to the others presented in state-of-the-art based on the LFW dataset for facial verification. A public CASIA-WebFace database was utilized in the training step of the introduced approach.