Research On Key Technologies And Algorithms Of Face Recognition

With the rapid development of computers and artificial intelligence,face recognition technology has made significant progress.At present,most face recognition systems can achieve excellent recognition rate under ideal conditions.Still the practical application finds that the recognition accuracy and eff-iciency of face recognition systems under non-ideal conditions are unsatisfactory.On the one hand,face images are often affected by a variety of dynamic factors,such as illumination and expression,which makes the development of face recognition technology diff-icult.On the other hand,with the arrival of the era of big data,the amount of face image data keeps increasing.In the face of massive face image data,the performance of traditional face recognition systems is low.In response to the above problems,this thesis mainly optimizes and improves the conventional face recognition system from three aspects:face alignment,face representation,and face classification.The research contents and related work of this thesis are as follows:(1)Accurate alignment of facial landmarks is the basic premise of a robust and practical face recognition system.Based on the ensemble of regression trees algorithm,this thesis proposes an improved ensemble of regression trees with NPD features.Firstly,the shape index feature based on NPD is introduced to solve the problem of weak anti-noise capability of original pixel difference feature in unconstrained scenarios and improve the feature expression capability of the algorithm model.Secondly,the similarity transformation matrix between estimated face shape and mean shape is found by using the Procrustes analysis,which provides a basis for the subsequent extraction of pixel features with different samples.Furthermore,the number of candidate features is reduced by the filtering feature selection method,and the computational efficiency of the model is improved.Finally,based on the idea of depth quadratic tree,a double threshold decision formula is designed,which makes up for the defect of insufficient feature division of single threshold regression and further improves the robustness of the model.Experimental results show that the method has a particular improvement in the accuracy of locating feature points in the unconstrained scene.(2)The performance of face recognition algorithms largely depends on the performance of face representation.Based on the Gabor quotient image model,this thesis proposes a face representation algorithm based on the weighted m-Gabor self-quotient image.Firstly,the traditional Gabor filter is transformed to obtain m-Gabor filter with rotation invariance and excellent curvature response,which solves the problem of incomplete illumination component extracted by a single-direction Gabor filter.Secondly,on the basis of m-Gabor filter,a smoothing weighting function is introduced to obtain the anisotropic weighted m-Gabor filter,which reduces the "halo"effect while preserving the dark features that are beneficial to recognition.Finally,the m-Gabor self-quotient image is obtained by using the idea of the Gabor quotient image algorithm to get the illumination invariable feature of a face image.Experimental results show that this method can effectively improve the face recognition rate in complex lighting scenes.(3)The time complexity of conventional face recognition algorithms will increase rapidly with the increase of the amount of face image data.In the context of big data,the conventional face recognition algorithm is inefficient and cannot meet the needs of users.To solve this problem,this thesis proposes a distributed k-nearest neighbor classification algorithm based on the clustering center point,which is suitable for a big data environment.Firstly,the classification model was built by using the sample cluster after Canopy clustering instead of the original sample data set,which significantly improved the classification speed of the algorithm.Secondly,the distributed computing framework is used to realize the parallelization of the optimized k-nearest neighbor algorithm,which improves the operating efficiency of the k-nearest neighbor algorithm in the big data scenario.Experimental results show that this method has a significant performance advantage in the big data scenario.There are 48 figures,12 tables,and 71 references in this thesis.