Low Resolution Face Recognition Algorithm Based On Multi-task Network

Face recognition has the advantages of strong interaction,no contact,and convenience.It is a hot topic in the field of pattern recognition.It is widely used in public security,intelligent education,pedestrian detection and other fields.However,in the actual scenario,since the hardware of the monitoring device is limited,the target pedestrian is often far away from the monitor equipment,and the acquired face image has a low-resolution,and it is difficult to achieve accurate identification requirements.Therefore,identifying low-resolution face images in actual scenes is the key to improving the performance of face recognition systems.With the development of artificial intelligence,the recognition method based on deep learning has become the mainstream algorithm in the field of face recognition.Because the face is non-rigid,it was affected by external imaging factors,changes in facial expressions,etc.Especially in low-resolution scenes,face recognition still has great technical challenges.In order to improve the recognition rate of low-resolution face,this paper will adopt a step-by-step strategy,first improve the face image quality through super-resolution(SR)technology,and then cascade the recognition network based on global separable convolution to build a multi-task network.The main contributions of this paper are as follows:1.Because the inconsistency of low-resolution samples and high-resolution sample information distribution,the reconstruction performance of traditional super-resolution algorithms is difficult to reach the practical application level.This paper designs a neural network based on separable convolution iterative projection.It is used to improve the reconstruction effect of low-resolution face images.The initial network extracts shallow features,and the projection unit network further reduces the gap between high and low-resolution detail information.The experimental results show that the proposed algorithm surpasses the compare algorithms in both the subjective quality and the objective score on the CASIA-Webface dataset.2.In order to learn the better decisive expression features in low-resolution face images,this paper proposes a face super-resolution algorithm for generative adversarial network based on feature enhancement.The algorithm learns features through residual networks with different depths.Dense connections enable interaction between different levels of information to learn hierarchical features.In addition,the weighted perceptual loss function of the feature domain and the pixel domain is used to optimize the network model to improve the performance of the algorithm.The algorithm is superior to the leading edge of image super-resolution algorithm both in subjective and objective quality.3.Aiming at the problem that the low-resolution face images observed in the real scene can not be directly recognized by cross different resolution.This paper proposes a low-resolution face recognition algorithm based on multi-task network.Firstly,the low-resolution image is recovered by the feature-enhanced generation network,and the feature with strong discriminative ability is acquired.Then,the network is cascaded recognition network,the feature extraction of the whole area of deep network by using global separable convolution,thereby improving the recognition rate.The experimental results demonstrate the effectiveness of the proposed multi-task network.