A comprehensive analysis of Local Binary Convolutional Neural Network for fast face recognition in surveillance video

Abstract

Recent research in convolutional neural network (CNN) has provided a variety of new architectures for deep learning. One interesting new architecture is the local binary convolutional neural network (LBCNN), which has shown to provide significant reduction in the number of parameters to be learned at training. In this paper, we study the influence of network parameters in the scenario of face recognition, comparing LBCNN against other famous networks available in the literature in terms of sensibility and processing time. In our study, we also propose a pre-processing step on images to increase the accuracy of the model, besides investigating its behaviour with noisy images. Our experiments are carried on the Chokepoint dataset, whose face subimages were collected from video frames under real-world surveillance conditions, including variations in terms of illumination, sharpness, pose, and misalignment due to automatic face detection. The conclusion is that by using the Laplacian step and a reduced amount of LBC modules, it is possible to train LBCNN more quickly and with improved accuracy. In addition, it was found that LBCNN is very sensitive to noise and better results can be achieved when noisy images are inserted in the training set.