Research On Robust Sparse Representation Face Recognition Algorithm Under Label Noise Condition

Machine learning is increasingly common in daily life,and the quality of training data plays a crucial role in the construction of learning models.However,the existence of label noise in practical application is almost inevitable,which will lead to a series of problems such as the increase of learning model complexity,the decrease of robustness and the reduction of generalization ability.Therefore,how to effectively classify in the presence of label noise has become a crucial issue.Through the study of this subject,we proposed the following three methods:First,multi-feature extraction is carried out for face samples in the dataset.The samples in the dataset are recognized by integrated learning model with multi-feature.Using the prediction results of integrated learning to build multi-label dataset for samples with inconsistent prediction label.And then the WSRC sparse representation model is established by using multi-label dataset.For the sparsity of represented coefficient in sparse representation,the classification with multi-label sparse representation can greatly improve the robustness against label noise and prevent the occurrence of over-filtering.Secondly,the local outier factor(LOF)detection algorithm was applied to remove the noise label samples from the training dataset.With the cleaned dateset applying KNN algorithm to correct noise labels,estalishing multi-label dateset to proceed sparse representation classification.In addition,in view of the errors caused by occlusion,light variety and expression variety in the dataset,we chose to use the robust supervised sparse representation classification(RSSR)model to perform multi-label sparse representation classification to achieve robust face recognition under label noise.Finally,applying the label noise filtering method based on the data distribution to filte training data,using cleaned dataset to correct noise dataset to establish multi-label dataset to perform sparse representation classification.Considering the real-time performance and the computational cost of the algorithm,we choose to apply KGSRSN algorithm to perform multi-label sparse representation classification,to achieve the improvement of computational cost and reinforce the robustness of face recognition model under the condition of label noise.