Face Detection Research Based On Deep Learning

With the development of the Internet, the data is explosioning.We need more intelligent methods to deal with the data. Among the data,video and picture data's process is demanding higher and higher. Face detection, as one of the most important technologies, has been paid much attention. In recent years, with the development of deep learning,computer vision related tasks have gradually used the method of deep learning. For the face detection task, people also began to explore the way by deep learning.In this paper, we first introduce several categories of face detection methods, and compare the advantages and disadvantages of each category.Then we introduce our cascade architecture of the face detection system which compared by low-level, middle-level and high-level three stages.This architecture can discriminate the candidate proposals from simple to difficult, so it can quickly filter out the non-face candidate proposals that are easy to determine, and focus on the candidate proposals that are difficult to identify. On the low-level and middle-level stage, there are not only classfication networks to discriminate the candidate proposals, but also the position correction networks to correct the position of the remaining candidate proposals. On the high-level stage, we apply Fast R-CNN[3] technique which from the target detection field to make final discrimination and correction for the remaining candidates. There are some strategies in each level of the system, one is the NMS algorithm which to remove the proposals with a Intersection-over-Union (IoU) ratio higher than a pre-set threshold to the selected proposal, another is using PReLU[29] activation function instead the commonly used ReLU activation function.After introducing the system architecture, we explain how to train the system. First, we introduce the training datasets AFLW[16] and Wider Face[28], and how we remove the pictures that are not suitable for training samples. Then, we enhance the data by several methods including blur and noise in order to increase the generalization of the system. Then, we explain the critical training process and parameters. Finally, different network deals with different tasks, so they have different training data.We introduce the training data of each network.After training all levels of the system, we tested the system on two authoritative datasets: FDDB[1] and AFW. On the FDDB dataset, when the number of error detection for the whole dataset is 2000, the detection rate is 91.87%, which is higher than the most methods in the last two years.On the AFW dataset, our average accuracy(AP) is 95.35%, also better than most of the methods in the last two years. We tested the average time of the system in VGA image with minimal face size of 40*40, it runs at 3.2 FPS on 1,2GHz CPU with a single thread and runs at 66.7 FPS on a NVIDIA K40 GPU.