Research On Face Detection Based On Deep Learning

Face detection has always been an important topic of computer vision.In the past few years,facial features have been widely used in access control,surveillance systems and other security applications.Face detection and identification is one of the most popular projects for the security department.The problems such as facial occlusion,illumination,extreme face rotation,low resolution,and scaling differences make face detection very difficult,and these problems are common in real-world applications.Many early face detectors have been able to detect front face images well,however,the performance of these face detectors cannot meet the requirements for a more challenging face detection set.This dissertation made intensive research and discussion on face detection under the deep learning framework.Based on the analysis and improvement of the existing deep convolution network model,a fast cascade convolution neural network model was proposed for face detection.Moreover,the proposed method was verified and analyzed on the WIDER FACE dataset.The main contents are as follows:1.By analyzing and testing the advantages and disadvantages of region proposal algorithm,Faster R-CNN,Joint Face Detection and Alignment using Multitask Cascaded Convolutional Networks,the region proposal algorithm and the region of interest detection network model were chosen for face detection,and a fast cascade convolution neural network model was designed.2.The training and testing strategy with convolution sharing was proposed in the cascade convolution network,which greatly reduces the scale of the cascade convolution neural network.According to the sharing of convolution layers,cascade convolution networks can achieve end-to-end optimization,which improves the performance of cascaded convolution networks.3.Since the proposal network and the RoI detection network share the convolution layers,the multilayer convolution information can be used in the proposal network and the RoI detection network to handle the smaller face area.This makes up the lack of Faster R-CNN in lower resolution and smaller face detection.Moreover,Leaky Re LU was used as the activation function to improve the detection performance of the deep network.4.Inspired by intuition in the human visual system,the fast cascade convolution network can be used to infer the face position with the body information and the face key point information.In this dissertation,the relative position of the body features was deduced from the position of the face candidate box in the convolution feature map.Then,the body information was merged with the facial information by Ro I pooling and L2 normalization.Moreover,the face key point information was added into the joint loss function so that the network has the ability to use the body information and the face key point information for face detection.Through the improvement of the existing algorithms,the proposed face detection algorithm with the fast cascade convolution neural network achieves a good detection performance on the challenging WIDER FACE data set.