Complex Scene Low Resolution Face Recognition And Its Application In Identity Recognition System

With the development of video surveillance technology,many video surveillance applications urgently need an unconstrained,long-distance effective face recognition technology.Usually in the real scenes such as schools,stations,streets,and shopping malls,etc.,the images that collected from these scenes suffer from poor imaging quality due to distance,light,angle,expression posture,etc.,and the resolution is low,and there are certain difficulties in recognition.Face images collected in complex scenes,most of which contain less features or features are difficult to extract.This thesis focuses on low-resolution face recognition using deep learning methods.The main work is as follows.In the aspect of face detection,we use a cascade classifier method based on Haar features.This method uses Ada Boost algorithm to quickly achieve face detection and extract the localized face images for face data collection.For image processing,firstly we expand the data set,and introduce several methods that for expanding the data set.Then,in the image preprocessing stage,we mainly used geometric normalization,grayscale,Gaussian filtering,and histogram equalization to preprocess the original image.An improved image super-resolution reconstruction method based on convolutional neural network is proposed.Compared with the original network,it has more nonlinearity and fewer parameters.There are four layers in this network,the first layer for corresponding to image feature extraction,second layer for feature nonlinear mapping and image reconstruction.The last two layers for completing the image reconstruction work,this method completes the reconstruction through autonomous learning.Finally,the method is compared with the traditional low-resolution reconstruction method.A face recognition method based on convolutional neural network is proposed.There are six layers in the network structure.Thanks to the network's local perception and parameter sharing mechanism technology,images can be directly input into the network and easily handling high-dimensional data without the need for manual extraction of features,thus avoiding features extraction and image data reconstruction in traditional face recognition methods,and experiments show that the recognition performance is superior to traditional identification methods.Finally,we design a set of low-resolution face recognition solutions composed of image super-resolution reconstruction,face location,image preprocessing face recognition and other modules,and successfully applied it to the identification system,it effectively improves the low-resolution face recognition difficulties.