Research And Implementation Of Low Resolution Face Recognition

Face recognition technology has the advantages of accessibility,simplicity,non-aggression.It has gradually become a hot topic at biological recognition technologies.With the continuous research of big data and artificial intelligence technology,the video surveillance system based on face recognition has been broadly applied to some fields such as traffic management,information security and criminal investigation.Since facial posture and expression changes are not constrained,the intensity change of light is uncontrolled,distances between collected persons,the surveillance camera is far and so on,the extraction of facial feature of collected person will be affected,which results in low recognition.Relevant scholars define this kind of problem which also beset with them as low-resolution face recognition.This paper focuses on low resolution facial problems including face feature hallucination,robust feature extraction and low resolution face recognition system The main research contents are as follows:(1)There is an existing low resolution face recognition algorithm on the basis of Gabor feature hallucination that adopts facial feature reconstruction to replace complicated facial reconstruction in the previous algorithm,which can be seen as a more efficient and accurate recognition algorithm,but it also has some drawbacks.It just simply reconstructs Gabor feature by a linear way and not take non-linear relationship between high facial feature and low facial feature into consideration.This thesis proposes an improved algorithm that adopt LLE(local linear embedding)algorithm on the basis of manifold learning;besides,in order to make full use of extracting facial feature information,the thesis continues to improve the existing subsequent projection space,then the following generalized canonical correlation analysis also be replaced by a multi-resolution generalized canonical correlation analysis proposed by this thesis.According to the comparison of experimental results at YALE and UMIST,an improved algorithm really enhances recognition at low resolution especially very low resolution by new feature reconstruction or new projection space.(2)The traditional single feature extraction algorithms are not robust to low resolution,and their recognition rate cannot meet the requirements.To improve recognition rate and enhance stability of this type algorithm at low resolution face,this thesis puts forward low resolution face recognition based on multi feature fusion which fuse local principal component analysis and local ternary pattern(LTP).Firstly,the features of low-resolution face are extracted by LTP operator and local PC A operator,then two facial features are divided into several blocks and histograms of each block are calculated;secondly,the histograms of each block on LTP and local PCA are linearly concatenated as well as each block is cascaded as a final facial feature,which its dimension reduced by equivalent mode.According to the comparison of experimental results,the proposed algorithm is more adaptive to changes of environment,moreover further improves the accuracy of face recognition at ORL and YALE.(3)This thesis adopts multi feature fusion and also use OpenCV vision library to design a low resolution face recognition system at QT software environment.According to a result of system testing,the low-resolution face algorithm proposed in the thesis can effectively identify low-resolution faces under surveillance cameras and meet better application requirements.