Low Resolution Facial Verification Technology And Its Application

Using the face images to detect somebody is needed in densely populated places, such as shopping malls, amusement parks, etc. Due to the distance between the pedestrian and camera and limited resolution camera, it always gets low resolution faces. Low resolution face image contains less information, together with different poses, illuminations, expressions, ages and occlusions making it difficult to extract effective features from the low face images.This thesis mainly study the low resolution face verification, which is to compare an known high resolution face image with a low resolution to verify the two faces belong to the same identity or not. The main content of this thesis composed of three parts:(1) This thesis uses the convolutional neural network to extract the facial features. Benefit from its deep architecture and large learning capacity, effective features for face recognition through hierarchical nonlinear mappings compared with artificial face features.(2) Imitating the thought of super-resolution reconstruction, using the convolutionnal neural network to process the low resolution face image makes the face image have more beneficial messages to identify. Most of the beneficial messages are high frequency information. Then use the convolution neural network to extract the facial features. Enhancing face feature's network and extracting face feature's network are in one convolutional neural network. This end to end network makes all parameters train together which is helpful to gain the global optimal solution and avoids the local optimum.(3) The high resolution reference face image and low-resolution test samples are mapped to a unified feature space to solve the feature dimension mismatch. In this thesis, two convolutional neural networks are used to extract features. Having the same configurations to extract feature can make sure the two networks extracting the same feature as much as possible. However, due to the different input image's resolution of two networks, the resolution of the feature maps in each layer is different. The network extracting the low resolution test face's feature has three more layers than the other one which is in order to enhance the low face feature. The network's fully connected layer is the final face feature extracted by the network. Make the number of fully connected neurons be the same to let the two networks to extract same dimension feature vector which means the two different resolution face images mapped to a unified feature space. The implementation of the convolutional neural network use caffe which belongs to University of California at Berkeley. After using two convolutional networks extracting the high resolution reference face and low resolution test face's feature, using the joint Bayesian to calculate the distance between the two faces to judge the pair of images belongs to the same identity or not.