Very Low-resolution Face Recognition Via Semi-coupled Discriminative Dictionary Learning

The traditional face recognition algorithms assume that the input face image has better resolution.Cased by the illumination condition,motion blur,the equipment noise and other factors,the target face images actually have small sizes,large noise and limited feature details.Low resolution(LR)face recognition problem is also a great challenge.In recent years,many very LR face recognition algorithms have been proposed.Most of these algorithms assume that there exists structural similarity between different resolution images.Based on the assumption,relational model can be built between the high resolution(HR)and LR images,then the transformation from LR features to HR features can be completed.However,these methods will encounter the following problems in practical application:(1)In practice the images' down-sampling will be affected by many uncertain factors,the process of degradation is not controllable.Strictly speaking,the relationship between the HR and LR is ‘many-to-one' correspondence.The assumption that the structure between different resolutions is consistent does not entirely fit in the very LR case;(2)At LR resolution case,the feature information of the image will be lost a lot,resulting in solving images representation in the sub-space will be unstable and exist certain deviation.In this paper,we will study the above two points: how to construct the learning model of manifold structure between HR and LR images and improve the discriminative ability of image features.In order to further clarify its promotion of recognition performance,this paper designs the three algorithms models based on above two points:(1)Face recognition algorithm based on fully-coupled dictionary learning.The algorithm is based on the uniformity of manifold structure.Firstly,the fully-coupled dictionaries are trained by sparse relational representation of different resolution images,then the association of geometric structures of different resolution in feature space can be further strengthened.For the face images in very LR,by preserving the manifold structure in the LR space to the HR space,the feature details of the HR image can be predicted.This method embodies the essence of manifold consistent learning;(2)Face recognition algorithm based on semi-coupled dictionary learning.This algorithm assumes that different resolution features are not completely coupled,and there is a certain mapping relationship between them.In the dictionary iterative process,dictionary structure can be constantly updated and we can obtain the mapping model of manifold structure between HR and LR images.For the LR face images,the manifold structures in LR space is projected into the HR space by the mapping model,which can reduces the structural error between the reconstructed high-resolution pictures and the original high-resolution images.The correctness of this conjecture is verified by experiments.(3)Face recognition algorithm based on semi-coupled discriminative dictionary learning.Due to the limitation of LR images' feature,the representation may appear deviation in LR space.In order to further enhance the image discriminative ability,The local constraints is introduced in images representation,and then optimization of the dictionary structure can be achieved by the semi-coupled dictionary learning.The experiment proves that the introduce of local constraints and the semi-coupled dictionary learning are helpful to enhance the discriminative ability of feature coefficients.