| In recent years,with the introduction of the “smart city”,a wide range of cities in China have been built the intelligent monitoring system.This system combined the technology of face recognition can meet a series of practical applications such as security,criminal investigation,personal attendance and so on.Face recognition technology development rapidly over the past forty years and a series of excellent algorithms were born,however,the captured samples tend to be low resolution and unconstrained facial posture,because the surveillance cameras is far from the pedestrian and unconstrained action of the target.Direct recognition of these samples surely can not able to play a good performance of the algorithms.Based on low resolution and unconstrained face samples,a single 3D face texture model is used for face frontalization.Meanwhile,we combine an improved recursive convolution neural network to get the high-quality frontal face image.Include specificly,the main research results of this paper are as follows:1)When the monitoring equipment collects pedestrian photos,the problem of image blurring and low resolution can be easy to occur because of the distance between the camera and pedestrians,bad weather and the movement of target.After analyzing the characteristics of several existing low-resolution face recognition algorithms,the structure of deeply-recursive convolution neural network is proposed to restore the high definition face image.In order to improve the network receptive field to obtain more high frequency information on the premise of ensuring the smooth training of the network,two extensions were proposed:recursive-supervision and skip-connection.Respectively,the perceptual losses and the mean square error as the training function of the network,the experimental results were analyzed to find although the perceptual losses network structure gets a low numerical PSNR and SSIM,but for the accuracy rate of face recognition algorithms it has a more obviously improve.Because of this a new idea of low resolution face image reconstruction algorithm has been found.2)For the problem that the monitored target's action is not fixed which leads to inconsistency of the collected samples,contrast to the existing features of multi-pose face recognition algorithms,a robust face-fitting algorithm is proposed.Firstly,the projection matrix between the input sample and 3D model is estimated,the standard front image can be restored by camera calibration technique and facial symmetry.Then we compute the differences of the local patches around each fixed facial feature points between the average face and test sample for occlusion detection.Finally,we combine the face symmetry strategy and the average face to fill the invisible region and occlusion region.We use the same 3D texture face model for all input samples,efficiently synthesizing frontal face images and improving the accuracy of face recognition algorithms.3)Samples of low resolution to frontal face,the accuracy of face recognition experiment is reduced to a certain extent due to the excessive noise introduced in the pixel complementation step.In order to improve this problem and so that the application of face-sharpening algorithm is more extensive,a model for image super-resolution and face frontalization was proposed.,and then the projection matrix between the input sample and 3D model is estimated for pixel position transformation.An improved recursive convolution neural network is used to make up the blank pixel and restore high-frequency facial detail information,because the recovery of the high-frequency facial detail information,the effect of noise is effectively reduced in the steps of final invisible region completion and occlusion estimation.The experimental results show that the model combining the two algorithms can effectively improve the recognition accuracy of low-resolution and unconstrained face images.Meanwhile,improve the practicability of face recognition algorithms in our daily life. |
PDF Full Text Request