Dictionary learning and sparse representation for the singular face recognition have made great breakthroughs,but when people hide their faces with sunglasses,scarf,and fake beards,lots of variations will be introduced,resulting in poor recognition performance.Another problems such as insufficient sample diversity,small amount available data will not make the training samples to represent the test samples well.The corrupted samples will also affect the discriminant representation to the test samples.So learning a more robust discriminant dictionary is essential.On the basis of summarizing dictionary learning algorithms at home and abroad,this paper proposes the following three improved algorithms:Firstly,traditional sparse coding with L1 or L2 norm to obtain sparse representation coefficient,which is not robust enough for face recognition with disguise,and only for identification,failing to fully explore the facial attribute.So the two-label singular face recognition algorithm based on iteratively reweighted sparse coding is proposed.The algorithm uses nonsingular samples as the discriminant dictionary,a weighted robust sparse coding model is proposed,the iteratively reweighted sparse coding algorithm is used to solve the problem,the weight obtained can describe the singular mapping of the face.Finally,weight is used to construct a singular dictionary for facial singularity classification based on it's physical meaning,and in the weight space the discriminant dictionary is used for identification.Secondly,the small samples size and the lack of diversity increase the difficulty of face recognition with disguise,so nuclear sparse representation algorithm with weber symmetric local graph structure(WSLGS)feature for face recognition is proposed.Firstly,the samples are divided into training set,gallery set and test set,training set contains non-gallery classes,and the training set does not include all classes of the gallery.Then,multi-scale WSLGS feature of samples are extracted,and projecting into the nuclear space,learning the discriminant mapping set from the training set,mapping the gallery to the set to constitute the based dictionary.A set with difference variables is constructed from the neutral and singular samples from the training set,several principal components are selected as singular model,doing the same mapping,forming a singular extended dictionary.Finally,the extended sparse representation is used for classification.The single sample from the gallery is used for classification.When a new class is added,only need to add the class into the gallery without retraining.Finally,when both training and test samples are corrupted,the low-rank representation technique can separate the singular pixels from the class-specific dictionary,but images from the same class are highly correlated,low rank constraints will reduce the diversity of atoms in the dictionary,and degrade the discriminant performance.Therefore,a low-rank double-dictionary learning algorithm based on multi-scale LBP feature is proposed.In the feature space,sufficient atoms of the low-rank class-specific sub-dictionary and the low-rank class-shared dictionary are guaranteed.Finally,the sparse representation and dense representation are used for classification. |