Pedestrian Face Detection And Recognition Algorithm Research Based On Surveillance Entrance Video

In recent years,the importance of video surveillance in social life has become increasingly prominent,and how to effectively process the data in different real scenes has brought challenges to researchers.With the significant changes that deep learning has brought to the computer vision field,algorithms based on neural networks have become research hotspots.This program aims to explore face detection and recognition algorithms in real surveillance scenarios,which analyzed the difficulties in lighting,blurred imaging,and unrestricted pedestrian movement.The algorithm was discussed and improved so that it can accurately complete face detection and recognition in real time in a specific environment based on deep learning.The main contributions of this article include:Aiming at the problem of insufficient local data in network training,a face image generation method based on Cycle GAN is proposed.Using the discriminator based on light CNN design and the generator based on Res-Unet.L1 pixel loss,identity loss,structural similarity loss(SSIM),and gradient penalty were added.Effectively improved the performances of style transfer and shape change tasks;With respect to the requirements of model miniaturization and local adaptation,a face detection algorithm based on improved MTCNN is proposed.Utilize the group convolution and the depthwise separable convolution to reduce the number of parameters,use log-cosh loss as the loss function of bounding boxes.The detection time was reduced with the combination of BGS algorithms.The network is trained by public and local data to improve its ability to detect faces in natural photo and the accuracy in the monitoring environment;For the requirement of network size and recognition efficiency,a recognition algorithm based on the Attention Model(AM)and the VGG network is proposed.Employ the AM to replace the high-level layers in VGG,combine the spatial and channel domain to generate the attention mask,use a one-dimensional bottleneck block to replace part of the FCN layers and extract different semantic information.A feature extraction loss function combining Triplet loss and Large Margin Cosine Loss is proposed,and the distance information in Euclidean space and angular space is captured at the same time.Which accelerated the convergence rate of the feature extraction model and effectively reduced the inner-class distance and increased the outer-class distance,improved the learning ability of the model.The recognition rate in the local data reaches 99.61% and on the LFW and YTF data sets were 99.06% and 95.12%,respectively.